Embarrassingly, I don't have one yet. I'm mostly still writing code one keystroke at a time. I've used Github Copilot to translate between languages and look things up, but haven't trusted it to write new code or fix anything.
Claude Fable is now good enough that I should be using it. The latest Claude version is said to be comparable to a mid-career graphics programmer, which is about what I need. Six months ago it wasn't good enough to use for such work, and probably would have made a hash of my code.
There's been progress.
But Anthropic's business practices are scary. They have a reputation for cancelling accounts for random reasons, offer rather bad terms, and have no customer support. I have a strong aversion
to services with a reputation for cancelling services arbitrarily. They will apparently cancel based on their system's guesses as to what you're working on. Sometimes they guess wrong. What I'm doing (metaverse clients) is so far out of the mainstream they may guess wrong.
I'm currently stalled signing up for Claude because their verification email system can't get to my inbox. (Some problem with their using Constant Contact, a spam system, to send transactional emails to my own domain, which, long ago, I had blocked at the Constant Contact end. I think. The verification email never touches the inbound mail server.) It's very important to Anthropic that you sign up with the "right" email address, and it may not be possible to change it, so using some disposable address is not a good idea.
Rolled my own for a Claude Max sub, basically tmux and an elaborate MCP for persistent semantic memory, Forgejo, etc. (https://github.com/ASIXicle/Coterie)
I sense more people will do so as you can fairly easily tailor any harness and MCP/LXC combo to suit the way you work. It's a bespoke future.
Subscriptions:
Personal: Claude Pro 20$/mo
Work: we have access to some Anthropic and OpenAI models and open-weight models, like GLM
Outside:
I have some open weight locally on my 64G ram M5 macbook, mostly for personal stuff that I don't want to send to cloud (or claude). Mostly using Gemma through oMLX, because it feels more street-smart than the more coding focused models, experimenting with Qwen-flash-next through Strata, because the streaming of experts really pushes what size of model I can run
Harness:
Claude code, mostly, I trust anthropic, and like the remote integration with their mobile app. Being able to say "prototype me this and that ui" for personal project and seeing it and testing it on my phone is awesome.
At work for the api-provided llms, I am trialing opencode and pi harness, but don't have much to say, we mostly used claude-code up til now.
For local I mostly use Hermes, not for any of the big features, just because I want a nice chat interface (I am mostly talking through my health problems and trying to get my appointments in order) that already has basic tooling integrations out of the box. I.e. getting OpenWebUI have search tool was pain. Adding i.e. mcp for my todolist or calendar to Hermes was smooth.
It used to be more interesting and elaborate, switching between different models. I used to use Cursor, then Codex, but since Opus 5.5 came out, I just use the Claude IDE. I got rid of all my skills and just set it to medium. Then I vaguely talk to it about very high-end things that I want, and it comes back with magic. Rinse and repeat.
I'd say the biggest trick is just to tell it what complete looks like. Be very specific about what you want it to come back with once it's done. That gives it a good idea what the goal is. It'll automatically create subagents and all that stuff. Although if you have money to burn and you're doing something large, you can tell it to just go wide or boil the ocean.
Also, make sure you're working on at least five different things. Otherwise, you're sitting around for one agent.
I can do 5 things concurrently for about a day max. 3 things concurrently all week and my weekends are spent untangling my brain. 3 things per day on Mondays and Tuesdays and 2 things per day concurrently on Wed-Fri and I feel ok but crave holidays.
I probably have some kind of pysccosis. I've had a personal $200 plan from one of the AI providers since March 2026. I'm never not making things now, I haven't rebooted my laptop for weeks because there is always an agent building something.
I cycle between 4-5 work projects and 2-3 personal projects. Every night I tell me agents I'll be going to bed and to not ask questions and keep going. I've always loved building things, but this feels more like an addiction than healthy sometimes.
They're not all done, and half the projects I bench, but this is where my personal projects live. https://wooten-labs.com
respect. this is how i've been working the last 4 months or so. it's a bit crazy making and i do think i'm enjoying things less because i'm preoccupied, but i sure do feel productive. lol.
Yep, Claude & Codex desktop are innovating extremely rapidly. Some of the features like code review are hidden under commands, but basically obsolete whole heaps of AI code review startups when you use them.
I just ask Claude code to do a code review of PR X and it spawns a sub-agent. No hidden commands or anything special.
If I hint it might be security sensitive code, it bumps the effort level and burns my usage.
But it works quite well.
The findings are often debatable or unreachable states but it does find stuff and improves the overall quality I find.
What's funny though is that it almost always inevitably finds something wrong, even if you just ran it in another window and fixed everything it found. There is really a point of diminishing returns.
Same journey for me, although I still use skills. Claude Code is fine; the remote-control feature is why I stick with it now. Generally anything in the cloud doesn't have a proper GPU that supports WebGPU, and as nice as SwiftShader is often you want to actually run things on a real machine. remote-control is by far the easiest way to do that from my phone.
Claude Desktop's UI seems pretty similar to Cursor now, almost. I'm still on Cursor but if I use it daily I often hit usage limits before the end of the month. Would you say using Claude directly is better bang for buck?
I use Claude Code Web almost exclusively for personal projects -- I’m guessing that’s pretty much the same interface but in the cloud -- and hardly ever run anything locally. Nicer UI than a terminal, real sandboxing, and it doesn’t tie up my computer. It’s just a remote agent that can make GitHub PRs.
INTERNAL:
Local Agent on PC via PI Coding agent
OS / kernel Arch Linux, 7.2.11-zen1
CPU AMD Ryzen 5 5500, 6 cores / 12 threads
GPU 2 × RTX 3060 12 GB (24 GB total), driver 615.78.08
RAM 62 GB DDR4, 3 GB swap
Model Flash-Next (IQ3_XXS, 75.8 GB)
Context: 65,536
Prefill - long prompt: 732 t/s
Tokens per second output: 53
Uses: Advanced embedded design, testing and coding
EXTERNAL:
Deepseek-V4-Flash on Opencode via Deepseek.com with prepaid Tokens
Usage last 30 days: $4.37 USD
Uses: Application building for Flash-Next above.
1. None
2. see above
3. Opencode and PI
4. no
5. no
https://github.com/oscscull/quarterdeck built my own harness. I had trouble with both claude and codex. In particular, managing multiple tasks in a kanban, syncing between my machines, making sure work is done on the right machine and can be triggered from either. Making sure I ping reviewers, meeting reminders, todos. Parking tickets til conditions are met or time elapses. Automated notifications when I have meetings. I have just kept adding features as I want them. This kind of software is so easy to make nowadays, I don't really think there's any sense in "launching" it for others, but I'm using it all the time. I highly recommend doing so (building your own personal work management software in 2026).
I guess this is also in some sense an agent orchestration tool, I don't use any of those.
Personally: I have a XTX7900 card which can run some decent models, eg, Qwen 27B and Gemma. For bigger stuff I have the Ollama Cloud subscription to access (mostly) GLM. I have a server with OpenWebUI, and for harnesses I use OpenCode, but I have been experimenting with pi.dev to create a workflow that takes the ideas from the Rational Unified Process (RUP) so that everything gets properly documented and reviewed by a human before getting to the implementation phase.
At work: Claude Code subscription, mostly with Opus.
I'm juggling multiple claude ultra subs, multiple codex $500 a month subs, antigravity for argon, plus API rate deepseek 4.1 flash, mercury 3, and abliterated large v2 for red teaming.
With just this message bus, I can coordinate AIs to work together at a normal pace (worktrees, subtasks etc...), but past a certain point the AIs can't keep up and I need to use a sprint card system I developed to coordinate work.
With this sprint system I'm able to work 200 cards wide (each card is around 30-60 minutes of LLM work in a session) across a fleet of 20 machines linked together with tailscale and ssh.
It's been very effective at getting high throughput, but afterwards I find that the resulting code can be a bit of a pastiche from different models working in isolation, thus a good pass over working directly with models to clean up and harden the code is necessary.
I find that using nova-bus to coordinate multiple model/harness review of codebase in very effective at increasing code quality. Each model sees things the others don't.
Is this amalgamation working on a single project? What harnesses do you use? How are you using mercury 3? How much and what kind of work are you sending to deepseek?
Can't work on a single project, Amdahl's law gets you then. Try to work on as many different unrelated projects as possible, so I can get maximum throughput.
I break work up into four loose tiers:
1. flash - deepseek flash 4.1 / luna / haiku
2. pro - mercury 3 / terra / sonnet
3. heavy - opus 5.5 / sol
4. frontier - fable 5.1 / astra / argon
The decision of how to classify each piece of work is done by the coordinator (typically Opus 5.5...). If a card fails at a particular tier, it's going to get escalated up tiers until it gets sorted.
Each card has multiple readers (code review). I find mercury is very good at reading quickly. Deepseek as well. I will be honest, I think for 99% of stuff I need to do Opus 5.5 @ medium is ideal. I just use the additional models for different points of view, especially the absolute frontier models fable, astra, argon at max effort for security audits and code review.
The sort of work I pass through this system, it's mostly library work for multiplayer games.
For example, my work on schema, my protobufs/flatbuffers language replacement for multiplayer games, parallelizes very nicely across 15 languages on the sprint system. https://github.com/mas-bandwidth/schema.
I did some prototype games, one thing that worked really well for me was having a deterministic replay system where the agents could drive the gamestate via recorded inputs. Kept me from having to do any playtesting or debugging in the main agentic loop.
Cool! I’d forgotten until now, but I had that exact idea back in the stone age, 10 or 15 years ago, and started working on it but got bogged down. It was clearly feasible but a huge amount of work, too much to justify the cost. So it’s a perfect job for coding agents now. Nice one!
I do you think now need to watch out for the opposite problem. These things become so easy to build that the temptation is just to go ahead and build. But unless you really do need them, all that new code just goes to waste. The cost to produce has become very low, but sometimes the value can end up being negative.
At work we have both Anthropic and Copilot subscriptions, plus Google AI as it’s included in our business plan anyway.
I have personally been using the Claude Code Desktop App and was happy with it until they made it harder to pick a branch for each worktree! Which is why I tried JetBrains experimental agent IDE, ThinkRail, and I love it! Super easy to start worktrees from any branch and the spec driven development experience they promote is great!
It also let me connect to Requesty for my hobby projects and with that I can easily try any model, from any provider. That gave me a much better appreciation of how models perform, how fast and how costly, thanks to Requesty’s super awesome dashboard which is a pleasure to use.
I was only using Opus 5.5 before as I wanted the highest quality, but using smaller models you can get much faster work, and the difference in price is dramatic. I have now settled on Haiku 5 for simpler things, it’s nearly free while being really smart still, and for big things I use Opus 5.5 still but also Google Argon 4, it’s incredible as well and it seems to write more human text than Opus.
We also have a framework that lets us launch Claude (we could use any harness but for now only Claude) in our Kubernetes cluster within a devcontainer. That lets us ship heavy work off to another machine so our laptop is not overloaded with agents. It works great but we are not open sourcing this framework yet so I cannot link for now. There are some alternatives out there but most of them are out to make big bucks instead of providing a small but powerful solution for this problem.
I build my own coding agent : Spurcode https://github.com/getspur/SpurCode
which enable to support me to orchestrate difference other coding agent : like Codex + claude/Goose ...etc
Also : give out product agent notebook workbench : https://getspur.dev/spurlab to make all coding agent need to explore/grouding before coding with notebook
I'm using a cli I wrote to orchestrate per-project VMs for sandboxing using libvirt with COW shared golden images to minimise the disk overhead. The VMs get "rebased" periodically onto new golden images for the rootfs, but keep a persistent home FS for longer. I'm going to clean-up the documentation and open source it at some point soon.
I was using oh-my-pi, but currently playing around with Hermes as an alternative.
Exclusively using open router at the moment, mostly on principle around ZDR, and freedom to experiment with harnesses - but financially I'll probably have to compromise on this soon. GLM, Kimi and DeepSeek are the main models I'm using, with some OpenAI usage.
Tmux for remote access and session durability. I'd prefer to find a setup where I could access the same session through a web interface and seamlessly switch back to TUI, ideally spawning sessions through a web hook or similar. I'm going to take a look at Hedr to see if this can improve this.
Probably the two major things I'm planning to do next are looking at adding a bifrost VM that will hold all API keys, and centralise management of MCP configuration, potentially use it's skill repo stuff as well.
I also want to start controlling egress more - probably iron proxy, though tbh I'd like a solution that was relatively liberal whilst still mitigating exfiltration and prompt injection risks. Maybe I can find a curated allow list to use as a base.
Finally, I make a lot of use of open-webui, but it's disjoint from the agentic setup currently - it'd be nice to find a better way to share information/research from it into the agentic sessions.
1) Codex $200 * 2, Claude Code $200, Muse Code $50, OpenCode $40. I run them in 6 machines (2 Linux, 2 MacOS, 2 Windows)
2) No. I try my best to use oauth
3) Claude Code + Custom Harness(set of skills) + Mod. This orchestrator (I call her as "Team Leader") orchestrates other machines with Tailscale.
- I've used Claude Code (Opus 5.5) by model satisfaction and mod.
4) None
5) Dots and Instinct. Main use case is research, code search and raise issue, PR reviews.
I use AI primarily through work, I don't pay for a personal AI sub.
I do have some OpenRouter credits but I mostly use them sparingly to try open weight models out for small personal projects.
I currently interact primarily with Claude Code, but most of my async workers are on Codex.
The orchestration part is custom. I talk to a "chief of staff" Opus instance to ingest various workstreams (Jira tickets, alerts, personal ideas, etc) and turn them into beads. It launches workers to tackle individual beads using Remuda, an agent launcher I wrote at work in 2025 and open sourced earlier this year: https://github.com/yendo-eng/remuda. I configure profiles in Remuda that launch containerized agents using my preferred harness and model. I update these profiles as the pareto frontier shifts. Currently it's GPT 6 Luna max for most tasks.
It monitors the workers and has them go through at least one round of agentic code review. It surfaces PRs for me to read once they've passed agent review. It then either ships them or hands them off to my coworkers for codebases where I'm not allowed to just merge PRs.
Around this workflow I had it write a small collection of skills and scripts to grease the wheels and make repetitive actions more reliable.
Since Remuda supports several terminal multiplexers, I've swapped out the backend 3 times: tmux -> herdr-tui -> herdr-core. There's experimental support for Rex but I'm waiting for it to mature a bit before I try and daily drive it.
It's not nearly as polished as some of the software factory setups and fancy GUI orchestrators I'm seeing, but it's light, hackable, and so far the bitter lesson has kept this working really well for me so I haven't felt much of a need to try anything else.
I have been trying out Muse for a while but it hasn't found a solid place in my life yet. Still too unreliable and too many privacy concerns to justify giving it a lot of access.
I have had a $20 codex subscription. I initially just used codex cloud but more recently have been using opencode with the open chamber ui over the top of it.
Codex usage limits have really gone down hill so I also now have a $10 opencode go subscription on which I pretty much just use DeepSeek 4.1 flash. I’ve found DS to be a great model - probably somewhere just below GPT 6.1 Sol but at a fraction of the price.
In terms of workflow, I discuss features with chat gpt (because it’s free and it has access to my project’s repo, read only view of prod logs etc, so can make informed decisions). Once we’ve agreed on how a feature should look I ask it to create a spike github issue. I then get another agent in opencode (usually DeepSeek flash or Luna) to break it up into implementation tickets.
When I’m ready to build the feature I can give open chamber a list of tickets to work through using the task plugin, so each ticket gets a fresh context. I try to build directly on the main branch and try to ensure we can always deploy to prod even if a feature is half finished (using feature flags etc to switch it off until it’s ready).
Once a feature has been implemented I get GPT Sol 6.1 to do exploratory / regression testing with a real browser against a test environment. This usually teases out a lot of edge cases I hadn’t considered and also gives me a chance to improve usability etc.
It often takes a few iterations of this (exploratory testing, fixing issues that are found, exploratory testing again…) before a feature is ready to release.
I can’t believe how far things have come since August 2025 when I first tried out codex.
I built a Qt GUI-based harness, Humbug, that runs on MacOS, Linux, and Windows (I have reasons to use all 3). Has only 3 runtime dependencies other than the Python standard library: https://github.com/m6r-ai/humbug
Usually have 3 or 4 instances running at a time and lets me use LLMs from 9 different providers (and not using their libraries either). The low dependency approach was slightly limiting early on but now makes it very easy to extend and makes it very easy for an LLM to integrate new features or optimize existing ones.
I typically use Claude Sonnet for document-related tasks, and currently DeepSeek 4.1 Flash is my current go to for coding (it's insanely cheap on ollama.com) and I find it's still fast and reliable at very large context lengths. Previously I used GLM 5.2 a lot and before that was Sonnet for most things.
Humbug lets one AI delegate tasks to another and it can choose different models/providers when it does, so works nicely for getting second opinions on things. I haven't yet found a need for an agent swarm though - more often the parallelism comes from working on multiple projects at the same time.
One of the biggest unlocks for me in the last few months is having AIs write ADRs to capture design thinking as a project progresses. It lets new AI sessions discover the intent behind the current work and surface any potential conflicts that might arise with new requests.
One fun feature is the AI can control pretty-much any UI surface the user can control too, and it's quite fun to watch it edit files (you can watch it scrolling around the file in an editor) or use a terminal. It's a nice visual way to see what's going on.
One of the biggest unlocks for me in the last few months is having AIs write ADRs to capture design thinking as a project progresses.
I find this works great as well. I try to do it on personal projects (without AI) with mixed success. Because it turns out I'm human. But when I started doing this on AI projects it made a huge difference. Features that I'd proposed and not implemented were picked up by the model and cross-cutting considerations were raised. Just insane how that raised overall output quality.
- Subscription: Google AI Ultra (testing for 1 month rn to see what i can get out of argon, after that probably going back to Pro package)
- Harness: Antigravity IDE : from what i have tested on commercial harnesses this one i liked the most so far
Regarding selfhosted solutions:
- Inference host: Im owning a 5090 and im running the models locally - mostly use qwen 3.8 27b q5 atm, its a bit "slower" even on a 5090 but it fits the vram i have available while also being quite good in terms of result quality
- Harness: I tested alot and ended up writing my own harness ( https://github.com/laughingmandev/loa ) since none of the ones i tested (yes i tested openhands, hermes, etc...) did work well for me. While my own harness still needs some enhancements, its especially made for long running / complex tasks while beeing hands off (not permanently prompt jokey'ing ) and sturdy enaugh to work well with selfhosted/local run inference.
I use the gemini4argon/google agent atm to work on a c++ simulation and image asset generation for a game im working on. The complexity of this stuff just is to big for a local qwen 3.8 to fully grasp (its not that good at c++ either). Local image diffusion also is not that great.
For other stuff of low to medium complexity im working with my own agent/harness/setup - especially if the data is anyhow security relevant since im not leaking anything outside. From writing scripts/tools to analyzing infrastructure or discussing architecture.
1. claude pro, ollama pro and opencode go
2. Openrouter for testing
3. I use VScode and claude code
4. I've actually built my own for agentic workflows - https://vectorstep.io
5. I use openclaw for my personal stuff, VectorStep for proper trusted workflows (and production workflows)
1) 2x Claude Max 20x + Chatgpt Pro
2) I dabble with open-code and hermes especially to try out the free models.
3) I use vanilla claude with an orchestrator on top
4) https://github.com/mhalstrom/maximizepm
5) I use Hermes.
- I use hermes as a "free" agent platforms as you can easily connect your chagpt sub and claude max plans come with api usage you can hook up. In my website https://app.toolscaled.com/docs/agents/how-agents-work/ instead of re-implementing agents and paying API costs you can hook up your automations to push to hermes, claude routines or chatgpt's equlvalent essentially getting agents for free which you can then incrementally convert them to workflows.
- Antigravity (AGY - Google Ultra $100/month), mostly gemeni flash but I benchmark Gemini Astra sometimes. My main driver because it’s pretty good at search compared to setting something up with DSH.
- DeepSeekHarness (DSH) Desktop using:
—- with pay as you go DeepSeek tokens, more expensive to use during my night/Chinese daytime, otherwise very cheap
—- oMLX on my MacBook driving mostly a Qwen 3.6 35b A3b MOE (mostly used for benchmarking a less capable model)
I mostly do custom subagent programming and can trigger subagents in AGY and DSH easily enough. DSH supports a Python coordinator orchestrating DSH subagents as conversations via RPC. AGY only supports a sub-agent orchestrating other sub-agents but it works well enough at this point. I plan on releasing my subagent framework soon (called autofab in pypi), which is basically a specification driven code development agent system (using existing agent harnesses rather than implementing my own).
I tried goose, OpenClaw, Claude, codex, etc…and they didn’t really work for what I needed. I might get a $100/month ChatGPT subscription to go with my Google one, primarily to get a different perspective and benchmark openAI models on codex (Claude doesn’t really support what I need it to, otherwise would go for that instead).
I use pi as a harness, and Paseo as a Mac UI (because I need a Ui).
I tried a bunch of models, but so far deepseek flash from October is just dirt cheap; and roughly on par with the frontier open weight models.
It’s still somewhat behind closed weight models (I only know Gemini Argon) in that it produces lots of text that is approximately correct —- frontier models seem to be very good at producing concise text that’s very correct.
I pay per token, using an EU hosting provider (melious.ai).
But I can fill in the gaps myself, so for me developing for fun, this is enough.
Currently I am using the regular Claude extension with VS Code, but one of my friends is working on a new harness called NodeCode that doesn't require regular compaction, so I'll be trying it out as soon as I can get some time to do it: https://nodecode.ai/
Nothing else, DS is smart enough to do everything I ask for and more, never hallucinates and never disappoints. Apps in Rust, Swift, Node, we even designed a new language better suited for LLMs, a compiler, and it runs beautifully. Haven't coded a single line since May 2026 and I will remember that date as the day AI took over my mundane job as programmer and promoted me to architect or director of my own apps
A custom plugin to ring a bell whenever a task is finished and a gong when a question needs an answer, that way I can play backgammon or mahjong in the meantime
+1 for DSH, it really should be getting more attention here. The only problem I have with it is getting search to work, but now DeepSeek has a solution, though you have to pay for it soon.
1. Claude (max) + Opencode Zen.
2. Tried some other models via OpenRouter but the pricing just doesn't justify usage sometimes.
3. Claude + OpenCode harness. Using Paseo sometimes to coordinate parallel tasks although I quickly run out of memory or grind to a halt doing this. Mentally I can't coordinate more than 3 or so coding tasks at once, and physically my Macbook pro suffers when going more than this. Convenient.
4. Forget swarms, I can't even handle 4 parallel tasks.
5. Tried, and these seem deeply overcomplicated, very AI-built, and ultimately not useful.
I built my experimental system and I'm using it. I'm genuinely very happy with it. If I had more compute and could run a real, full 30B model locally, I'd be absolutely thrilled.
1) I use 2x ChatGPT Pro 200, and a Claude Max 200.
2) I use other models via Open Router, but ChatGPT beats them in value (at the moment) thanks to Tibo's resets.
3) I mostly make Claude coordinate a group of Codex agents. They've set up a communication channel between them, and Claude acts as the coordinator and a second verifier. This saves me time because it's easier to talk to one agent than a group of agents. Codex parent session does only "comms" with Claude, and all the other agents run as sub agents.
4) For non coding usecases, I use https://www.ctlflow.com. This lets me run dozens of agents on tasks like equity analysis. The information a swarm of agents can surface from a large data store can be quite surprising.
5) I have tested OpenClaw, but now I just use ctlflow.com or claude + codex.
> 4) For non coding usecases, I use https://www.ctlflow.com. This lets me run dozens of agents on tasks like equity analysis. The information a swarm of agents can surface from a large data store can be quite surprising.
Can you tell me more about #4 regarding what your heuristics, methods, and use cases are for equity analysis?
I ask because currently I am building a narrow use case Python app using Claude specifically to filter equities that may have an asymmetric potential to buy LEAP options exceeding 1-2+ years.
I am not a finance/quant bro, but I do want to get better at leveraging AI to be a better investor.
> 3) I mostly make Claude coordinate a group of Codex agents.
Same here. Are you just using a skill or something else. Beyond direct CLI calls to each other, I’m using contextify to share past sessions between them.
I'm personally primarily using Claude Opus 5.5 via Claude Code on a Pro subscription. In my timezone I seem to get an incredible amount of usage equivalent to about 2 full working days per week, letting it do all the coding but focusing it frequently. It blows me away how good it is and how much usage they are giving us.
At work I'm afforded a Claude subscription that is 5x that much usage and I simply cannot use even half. I'm not coding at all anymore at work, just driving Claude. It only needs a bit of direction.
I also use OpenRouter with OpenCode, and a homegrown harness I wrote initially without AI help but later on with AI help. But I haven't used this setup for a while. I swapped models based on the churn as reported by artificialanalysis.ai but they don't release as much data via their free API anymore.
I also can run Qwen 27B A3B on my NVIDIA 5060 Ti 16GB but I don't bother with that anymore either.
I don't use orchestration or swarm. Occasionally Claude spawns a few sub-agents as it needs more data and is preserving it's context window.
1) Claude Max 5x, ChatGPT Plus (for minimal usage of Codex, mostly fired as a Pre/Post PR hook for Claude Code to get reviews from GPT models locally
2) None for now.
3) Claude Code CLI, but not always in a terminal I'm sitting in. I talk to it on Telegram and Slack for different projects.
4) Something custom that grew into a project: Spool (https://github.com/enes-alatas/spool). It runs long-lived Claude Code loops, each with a name and a mission. Each one wakes on its own timer or when someone messages it, works, and sleeps, and the session resumes with its context. Each loop gets its own container with an egress allowlist, and they message each other and me in a Telegram or Slack group. A web control room shows what every loop is doing and what it costs.
I use free models in opencode, mainly DeepSeek, but it turns out my laptop can run Qwen3.8 flash next at 60 tok/s via Strata. So I mostly use Qwen now in a customized version of pi. But when I need the GPU to be free, I have to use free models in opencode. I don't use any orchestration - free models don't have high enough limits for that, and local model has only 1 conversation cache slot, more won't fit to my RAM
I have a CLI based hybrid setup on my MBP (M5 48GB) and it has proven to be very effective.
1. I have a $20 Codex subscription that is more than enough for my work. I use a 75-35 model where I used higher models (Sol/Astra for 35% of the heavier tasks and lower models 75% for other tasks. Sometimes I also use Free OpenRouter models for minor tasks.
2. I use many local models like Qwen3.8-27b/Gemma-4-26b and some 9B models. I also heavily use voice/vision models like Kokoro, QwenTTS for marketing and product videos.
3. Pi exclusively, which I have tweaked for my hybrid setup.
4. I have my own custom agent orchestration that routes between Codex and local models based on the task.
5. I used OpenClaw for a bespoke market research task once, it worked flawlessly.
I have both platform.deepseek.com (hooray off-peak pricing) and the $100/m Claude subscription and boy do I find myself mostly using DS-flash pretty much every time.
There's something to be said for the speed with which it gets word done - I take fewer "it's compiling" breaks, and lose my train of thought less frequently. What's crazy is that the speedy work it does is pretty much always correct - which is a reason why I don't feel the need to spin up claude-code and "ask the expert".
This speed & good-work combo is the DS signature IMO
For my harness, Deepseek Harness + DS-4-1-flash combo is almost too good to be true
Harness: Claude Code plugin for VS Code. At some point I guess I'll switch to just the claude code app but tbh this is super ergonomic for me.
No agent orchestration besides literally just letting claude spin up agents for me.
No broader systems.
And I'm _massively_ productive with the setup I describe above. I say this not to gloat, I'm sure others have great methods but I do say this in case anyone is feeling overwhelmed by the options out there. My 2 cents is don't overthink it.
For example, say you want to search for bugs in 100 code files. Ask Claude to “use subagents to look for bugs across these files” and it will likely spin up 5 agents, each taking a different section of the codebase. Theoretically in the limit case this is 5x faster than Claude doing it without subagents.
Any kind of problem shaped like map reduce can be accelerated this way: Looking through your banking records, searching through a database, asking Claude to find that one weird file you remember from three years ago but forgot where you saved it, etc.
Just know that there’s no quicker way to hit your limits than to use subagents. So use them if it’s paid for by your work or whenever you need the 5x speed boost.
You can tell it to do so, but what you'll find is it knows to do it when it needs to. This is another case of "don't overthink it". Just let your agent decide when spinning up sub-agents makes sense.
If you focus too much on optimization, and you do so in the wrong ways, you'll actually find you saved no time/money/tokens vs just letting it decide when to do it. This applies as much to Agentic engineering as it did with programming pre-AI.
tl;dr it will literally do it for you, and it'll tell you when it does so
2) Local LLMs on an M1 Max (Unsloth, LM Studio, occasionally direct llama.cpp). Muse Glimmer (because it is good at what it is optimised for: interactive coding sessions). Qwen 3.8 27B occasionally and a couple of other things for fun (an LFM MoE, experimenting with Nemotron Bolt 74B-9B, enjoying the really super Gemma 4 12B QAT when working on creative photography research or non-code planning).
No openrouter (yet) but that will definitely happen.
3) Currently focusing my attentions on Pi via Paseo (and trying Pendant). Pi is just much better in smaller context sizes, and it is the right fit if you are keen to learn from first principles.
I tried OpenCode first. I couldn’t quite get on with it and what ultimately made me switch was trying to understand its terribly messy source code.
4, 5) I am mostly using AI as a coding sidekick so I don’t really have much need for orchestration yet. I did some experiments with Osaurus (really quite impressive) and I actually think that is almost enough by itself, except that I want a little more control over the sandbox container.
I feel that my usage pattern is still settling, because I am not keen to surrender all of my coding to the machine. But I am trying more purely “directed” coding to learn the process, and finally beginning to get a feel for how I will work with that.
t3code on my mac and omarchy, connected to my VMs with tailscale, that runs inside my own proxmox cluster.
I get to separate my work, have a lot of compute, manage it from anywhere, and also have my agents go bananas in their VMs.
I also run 1x Claude Ultra, 1x Claude Max, 1x Grok with X Premium, and 1x OpenAI sub.
I think the biggest leverage move for me was to get the VMs. Now I get to group my work better, and give my agents a lot more power by running more freely.
t3code has been a solid boost as well. It did hurt really bad tho, because I have been a terminal nerd for a long time. But the speed I've been able to achieve across a lot of projects has been crazy with t3code.
thb for me it's been the workflow around it. Things like:
- Pulling up my phone, dictating a quick thing and sending it of to work on one of my servers.
- Shooting things off to claude code, codex, grok, so fast. So well integrated.
- Hitting snooze on a session for next week
- hiding sessions in snoozed and working, so I can focus on what needs me now.
- Dragging filed straight into the chat
- Quickly looking at the diff on the side and go through files with the "viewed" button as in github.
- ctrl+j to get a terminal quickly to do something fast
- ... So many more but this is top of mind.
I just get so much more done, faster, with less context fatigue than before because of how it's laid out.
1. Claude Ultra. I've never had multiple subscriptions from the same provider, though that has looked more attractive lately as I
2. Yes: I often have personal tasks that Claude refuses to do, so I have to use Deepseek-v4.1-flash or GLM-5.3. I typically stay with OpenRouter, but I've found a discount provider that offers a 2/3 discount and have migrated to them. The catch is the data is probably sold, but given that it has saved me $300 this past month for non-sensitive work (game decompilation), the tradeoff is worth it.
3. Claude Code for Claude. I used to use OpenCode for my API usage, but there are too many issues which have gone unfixed and too many weird issues that I became tired of working around, so I switched to Oh my Pi. I'm not a fan of it either, so before my next big run with open weights models, I will probably rewire Claude Code to use these Chinese models.
4. Claude Code's subagent feature and workflow (subagents, but queuing up >200, and running 8 at a time) are sufficient for my needs. I had Claude build some bespoke workflow orchestration tool for managing many OpenCode agent instances, but it is a single shell or Python script that is essentially the equivalent of Claude Code Workflow.
I have a $1k/month subscription from cursor for work. depends on the task priority, complexity I select between opus 5.5 and auto (grok). I'm an escalations engineer
pay-as-you go on openrouter. use latest DeepSeek or Gemini.
I created my own minimalist/task based harness that checks out a jira issue, downloads attachments, runs sandbox docker containers with configurable gh/local repo as volume mounts depending on jira labels, can load agent instructions from local/remote depending on jira type (triage-bug, dev task, reproduce-bug, security) has scoped/temporary access to targets (e.g. dev k8s cluster, aws/gcp/azure account, ssh, splunk). dynamic secrets loaded from external secret management system/env vars.
since a lot of the corp repos are littered with skills, rules and other crap that bloated the context, I also set enabled a way to ignore repo/org agent instructions.
I have 2 repos for storing my personal/corp agent instructions and I 2 separate harness configurations.
My current setup is Herdr + oh my pi + pi. Then I mostly use oh my pi's "prewalk" feature with Opus/Sonnet as planner and Deepseek Flash 4.1 as executor. I use OpenRouter to access the models.
This setup is working really well for me and provides a good balance of cost and efficiency.
I use a custom Daemon. I tried OpenClaw and Letta and it was like emailing your AI. I have a 64gb Mac m1 Max and the largest model I could run in openclaw was a 27b because its not built for local AI, it's built for API. I wanted something that you could chat to, collaborate with, build, evalutate the code, make corrections, build, repeat. I couldn't find anything that didn't come with massive latency so I just made my own. Now I'm running a 125B A3B model at 30 tok/s using SSD Streaming and my agent is instant with no latency and can do everything I ask. I use Aider for programming, exa for searching.. I'm happy with it.
I'm locally hosting variants of Qwen and Gemma 4. I've also tried inference services such as baseten and novita, but wasn't satisfied with inconsistency in performance and pricing.
For software dev, I primarily use my own harness/orchestrator (that I shared recently), and sometimes Pi. I switched away from Cline + VS Code primarily because code is being generated so fast that I needed to optimize for review.
My harness is modeled similar to my day job: dev team works on stories, tracked on a kanban, and I'm a code-review gate. For example I scrutinize control surfaces, but instead of reviewing the implementation: I skim it and ask questions.
Every now and then I check to see if other harnesses have what I'm looking for, but none seem to have a combination of:
- Strong sandbox
- Particular UX (for mobile, code-review, basic IDE, and task management)
- Code checkpoints
- Great compatibility with local models
1) Kagi Ultimate. I don't use any subscriptions for agentic coding, just search. I evaluated using Perplexity Pro for a while, paying for both subscriptions at the same time for a few months. I find it hard to describe the experience of using Perplexity Pro beyond making me feel physically ill when I used it.
2) I primarily use Qwen 3.8 flash q2 on an R9700 with 4x256k context. I also have some Openrouter credits as an escape hatch if I need it (usually to work on my local AI configuration, sometimes to speed up certain tasks).
3) Pi[1] with a suite of extensions and context files that I have built out myself for the most part, with a few external extensions. One of the main external tools that I really like for coding is openspec[2]. I switched to Pi from Codex because I liked the minimalist philosophy behind it, and because at the time I was using a 7900 XTX (which is now my primary desktop gpu, and I also use for auxiliary services like whisper and decision models), and I was having a hard time getting 100k context with any reasonably useful model. The default prompts of Codex ate into that a lot, and Pi was a much better match for the resources I had.
4) I use Herdr (tui)[3] to manage Pi agents. All agent to agent interactions that I do are ad hoc. I'll do things like tell an agent to spin up work in several other Herdr panes that I have opened with Pi in them, or I will do things like tell an agent to inspect the work of another agent and make sure that it continues to stay on task.
1. I've been using the $100 github copilot tier but am looking for something cheaper, as my usage is waning somewhat. I like to hop between models for comparison, but openrouter always seems to end up being way more expensive for the flagship models. I also keep a GPT Plus subscription for voice, until someone has a comparable offering
2. No outside models yet. Working on a setup for that. Sometimes I throw some money at openrouter to check out the mega-cheap options.
3. At home I really like goose (TUI), but have been messing with opencode (TUI mode)lately which is also nice. At work I use Zed with a corporate claude subscription and it's a good setup.
4) Not really. I tried some of the opencode swarm plugins awhile back but they didn't seem to work right. In general I don't really have the energy for messsing with a new plugin ecosystem. I also like to work in a _tight_ loop, conversationally as things are being built, steering all the decisions. I have next to no desire to hand a larger task to a "team of AIs" to replace my input with more AI.
5) No
I drive a lot and spend a lot of time talking to GPT voice workshopping ideas, talking at a conceptual level, looking up things I'm curious about, and discussing design and problems I'm having, sometimes even work problems. I use goose for all kinds of stuff, mostly infrastructure and debugging for now but I've written a couple programs with it as well.
I do not like terminal tui much as I think it's a poor fit for a coding agent.
But I do use claude code tui because I have a sub: it is a good model and I like its aesthetics. If I could use the sub in my own harness I would. It's also a foil to GPT's models that overengineer everything.
I have a Codex subscription that I use in my own harness, dreamcoder.ai, which has a tiling window manager and all the other mod cons you'd expect a frontier harness to have.
In addition to that, I run Qwen 3.8 27B NVFP4 with ninfer on my 5090. It's a great model for local development. I also use that in dreamcoder.
I also do orchestrator like work that way. The key difference is I'll have a master chat and let it delegate work as needed to sub agents.
I have skills, MCP etc for things that need it, and custom-designed sub agents for a couple of very specific, niche tasks. The rest use generic delegator, explorer, worker sub agents that just take normal instructions.
Can you not? I built a custom harness that uses my Claude subscription. You can pass -p I think is the flag, to feed it a prompt. And it’s been a little while so I don’t fully remember how but you can also have it print the output to stdout. I even pass it custom tools as args and it works with the cache, and you can even append messages to conversations
1) Opencode GO 10euro (claude max at work payed by empl)
2) Online opencode
3) Opencode harness is better at keeping your costs low but claude code on ultra code is simillar to agent orchestrator
4) Agent orchestrator
5) Ive tried hermes agent but my little server cant hold it :(
1. Claude Max with Opus 5.5 and the ChatGPT $20 plan, I have a few thousand bucks of credits since they reduced the Codex plan so I’m using it for all the bulk data tasks that it would be a waste to use Claude for.
2. ChatGPT Image still looks really bad to me so I use Gemini Image. I tried Argon but it really didn’t work for me so I just use the image model at API rates.
3. Just the default Claude/ChatGPT desktop apps. I used the command line tools for a while but the desktop ones are much easier to use when working with images, and you get the usage down the bottom.
4. No, I think this is all a big gimmick. I barely use subagents. It makes sense to have multiple people at a company with different roles due to their different capabilities and experiences - but when the models are identical clones then why not just have one do the whole job? I only use subagents for bulk data tasks because otherwise they get distracted chitchatting with each other.
5. No, I haven’t tried them yet. What’s the benefit over ChatGPT or Claude?
If you use claude code on ultracode it already spawns subagents by default now, i think agent orchestrator has the only benefit of ofloading some tokens and making the subagents to not get distracted by the bigger picture but i cant see any use case of the broader orchestration on a project that is already solid and you just push features and debug, a single session give you more ownership of what is build
Yeah, the ability of the sessions to message each other turns it into a mess. They all start messaging each other with unrelated tasks and you end up not knowing what each thread is meant to be working on.
I notice that single-threaded agents get stuck with things sometimes like writing some tool they need inefficiently so that it takes 3 hours when it could be 3 minutes (meaning that a task ends up taking days instead of hours, unless I intervene and tell it to make a fast tool instead) but subagents just compound this as they burn through tokens without realising the problem.
How about, what sandboxing are you using to prevent the LLM from doing something you don’t want? I’m using nono.sh and have restricted network and file access. I use nono’s proxy so the LLM never gets access to api keys. I create a separate api key for each service specifically for LLM usage, and I restrict those keys as much as possible
I built something which auto-containerizes the agents for you (gvisor isolation, with credentials injected via proxy and url allowlist), but still allows orchestration between the agents, and a desktop app as well: https://github.com/bsklaroff/yaac
The wrapper also keeps history, memory, etc. split by project folder, so agents working in different projects don't know each others' history. If anything, `--resume` is a split second easier.
I have been working on a shared folder that sync between agents (https://www.filestash.app/agents/). The control plane [1] handles pretty much all storages protocols / platform (S3, SFTP, NFS< SMB, ....), you add the controls / permission boundaries and everything is auditable + traces, and the sync client is at the edge [2] to sync the data it actually has access.
Tried nono.sh but it seemed cumbersome. I think it lets the whole agent run in the sandbox, so whenever you need to change something, you have to restart the whole thing, if I remember right.
Also tried Docker Sandboxes which worked fine during the try out but my trust in Docker products with regards to security is low in general, so I'd wanted to wait a bit whether problems show up.
Then used Codex for the built-in sandbox but it's buggy like hell, documented features just do not work properly. Nobody there every seemed to actually try using the stuff they are documenting as working.
Ended up with 3code right now, which also has built-in configurable sandboxing.
1) Home - Claude $20, Gemini web, ChatGPT web. Work - Claude $200.
2) No.
3) Claude Desktop (home/work). Cut and paste when using Gemini or ChatGPT
4) No
5) No.
Claude Code Desktop on Windows: I like it. Simple UI that I started using about 2-3 months ago and have grown with it. Multiple sessions, virtual groups for sessions, etc. I can see my manual workload of managing things is getting larger so I want better project tooling. Some of the default selections tend to stick with what you selected first and I dont like those aspects (new sessions seems to default to your original/first session folder - so we need to clean that up or create some built-in /new-session skills, etc)
However, what I really want now is voice integration without having to click on the speaker icon and talk to it (or a way to keep the speaker icon depressed) . I want the ability to get off my chair and away from my screen and talk to it. This part is easy, I think, and would be a good start. To close the loop I would like to hear the feedback. However... getting the responses to have a 'conversation modality' instead of a 'verbose text modality/Claude speak output' may be a bigger challenge. The workflow will never be 100% display free. But if we can get the audio aspects worked out, projection could be to a pair of smart glasses (XReal, Apple, Meta, Google) - activated only when needed (read something or see a UI design, etc) and then pushed aside.
Anyone one have a minimal keyboard/display approach working for them?
1) Only ChatGPT Plus at the moment cause I'm between jobs and couldn't justify all the subs.
Grok gave me three months at $30 when I tried to cancel so there's that too.
I wanted to really keep Claude but ChatGPT voice mode is too good tbh and no other comes close. I wish I could use Opus 5.5 for coding.
2) OpenRouter PAYG, mostly for experimenting with Chinese models like Deepseek and GLM, usually for stuff that Codex or Grok refuses to do.
Had been looking into NEAR and Venice for completely private information processing (like emails) but haven't set them up yet.
3) Codex, and OpenCode as harnesses basically.
4) I really wish I could decide on some agent orchestrator without trying them out for weeks. I like my tools highly powerful and trying to find recommendation for this online is hopeless because it's full of opinion from people who never used a terminal before.
5) Grok Bot for running experiments with supplements, calory tracking etc.. Tried to set up Hermes a couple of times but hated it every time. Just gave Rakazo a try yesterday and still hating it.
Gosh, I dream about the time when you could look at a github repo and make a decent guess about the quality of the software.
I use Paseo for mobile / desktop UI, and as an agent runtime for some automation in a dedicated Hetzner box.
I try to keep the harness setup simple and portable, I use Claude Code or Codex with their subscriptions, I invest a lot of time in my skills which are shared between them.
Deepseek/Deepseek Harness with about 30 plugins and a lot of customization. I tried IDE integrations with Jetbrains for a few months and hated it. Now I enjoy juggling 3-4 projects at once. Costs maybe $20/week. A lot of jobs you can work at for a long time with little cost because there's a 99% cache hit rate.
Sometimes I despatch it on open-ended tasks that I'd never take on otherwise; for example I have a little fleet of 10 agents digging up medieval Japanese Buddhist parables, translating them from scratch, and feeding them into a custom interface optimized for language study.
Most of the time I give it a moderately detailed description of the scope and key/unique features I want in a new project, and it banks on work from previous projects to know what sort of deliverables I want, track and review its own mistakes etc., then I steer it from a rough prototype into something more polished. I find it equally useful for coding/application development and research/synthesis tasks.
1) Subscriptions - None
Using OpenCode BigPickle
2) Models - OpenCode BigPickle and deepseek v1
3) Harness - OpenCode as primary , but started liking Pi recently
because of all the customizations
4) None
5) Played around with OpenClaw for a really long time , now been exploring
the Hermes Agent
I use T3 Code as my front-end for all my coding agents on all my personal computers and my VPS.
For the actual agents I use Claude Code for my Claude subscription, Codex for my OpenAI subscription (I usually only keep one of those at a time), and pi for my OpenCode Go subscription.
I mostly use Sonnet 5.5 medium and Sol 6.1 medium for everything. But lately I have been using DeepSeek 4.1 Flash a lot and find it very capable.
At work I use T3 Code as well, but it isn't connected to my personal machines or subscriptions and use Github Copilot in T3 Code via ACP. Model choice is similar, but for more involved tasks for architecture analysis or design, big features that touch a lot of interconnected subsystems, etc. I'll turn up reasoning effort to high or use Opus 5.5 on medium or high. But when actually implementing, I try to break up my tasks into smaller pieces and the models usually do well even without reasoning turned up.
Most of my work goes into rewriting prose in reports, commit messages, PR summaries, etc. because all models still struggle with that. (Claude models less than GPT ones in my experience, but still ...)
I just recently started using T3 Code and it's strongest points are having a really good workflow, interconnecting all agent harnesses so you can easily orchestrate across them and they can easily communicate, and truly phenomenal remote sessions. All while having a singular UI/UX experience which I actually really appreciate.
It is probably weakest at customizability and making sure you have parity in all your agent harnesses with skills/etc, granted I just set an agent to ensure that so the point is a bit moot.
Overall I have really been enjoying the inbox style workflow and it has been a joy to work with. It jives with how I like to work across streams.
It has no weak points in my experience (compared to other harnesses). It is so good that I just can't use anything else anymore without feeling frustrated.
Theo of T3 Code has made a recent video about its features.
A few examples of why I like it:
- Not having to use multiple agents (you can switch them mid-conversation with automatic handoff if you want)
- The way it handles environments (I can start a new session on any environment (machine) with two clicks (new thread -> pick project on any environment)
- I can ask for visual responses (uses a T3 Code tool and renders a html page inside chat with cards, charts, etc. that follow the currently selected color theme (even if you change the theme after the response has been generated)
- MCP tools like browser, webfetch, etc. so I don't have to set them up for every agent
- Thread handling: Snooze, Settle, Archive, Delete. It's the only tool where I don't get overwhelmed when working on multiple Threads/Projects at the same time (I am diagnosed with ADHD).
- Provider usage overview for all registered providers, showing pacing as well
- Great citation: when you make a selection in chat, a pill appears and you can create quotation elements that get added to that, can easily be edited or removed (the original reason why I even tried T3 Code in the first place, because it makes responding to long responses incredibly joyful instead of disorienting). I like it so much that I added the same mechanism to my own personal assistant/agent that I host on my VPS.
- Great diffing view where you can pick worktree diffs, or any diff from a specific turn in the current thread
- Has it's own subagent orchestration that's very pleasant to use
- Many more little things that remove a lot of friction and just feel right for me.
It's just baffling to me how it is better in every regard than every other harness that I've ever tried (Claude Desktop, Github Copilot App, OpenCode Desktop, all the CLIs, Integrated Agents in VSCode and Zed, etc.)
I will still start up a pi on the cli if I just have a quick question because T3 Code needs a bit to start up and connect to all providers/environments. This is my only point of friction: It's a bit heavy on resources (because it does a lot, not because it's slow. I just happen to use it on very old hardware and on one device that is very slowed down due to heavy-handed security scanning). But other than that, I can't imagine using anything else right now.
I don't know if I've ever been this exciting about a piece of software. I was very sceptical at first, but quickly fell in love with it.
1) Claude Pro $20/mo (I'm retired now and just playing on personal projects)
2) I have API usage via Together for a number of toys I have built for my kid, I probably could refactor this but it predates getting a Claude Pro subscription.
3) Claude Code letting it do the work, pure vibes and dropping to vstudio as needed.
4) No
5) Muse, I have it doing things like basic research, waiting on ticket windows and just being a personal assistant. It also does a really good job at scraping churning forums and blogs and has made a few thousand dollars in the last month just on bank and brokerage bonuses. It does all the record keeping and I just sanity check the dates, hold periods and do the account opening / transfer work. I'm eager to see Anthropic's take on this space and hope it uses the same Chrome integration. Muse's biggest flaw is that it runs it's browser actions from a datacenter instead of a local host and a real browser.
Being retired in the age of AI is the best, all the time in the world to play with what really interests us plus an incredible assistant to squeeze every optimization path to productivity. Apps developed in minutes instead of months is like adding another century to our lives. And we are just starting...
My manager described this as incomprehensible lol but it’s modeled on conductor but uses Claude code terminal and mods and glues a bunch of open source stuff with a cli I had Claude write. Herdr for sessions, nvim for code review / searching, cli invoked by /slash commands. Beads cli for task tracking. With multi repo support since I end up doing a lot at work across repos at work for a “feature” which is how I delineate chats now. And I’ve got a nice picker for swapping sessions so screen real estate isn’t wasted. Many incremental improvements the sum of which are compelling ime
Phasing out my usage of cloud based LLMs/harnesses.
Focus is pi + self hosted GLM Flash.
Pi is lightweight and easily extensible and creates minimal extra noise. the output I get is very close to what I want to see.
my workflow is focused on clear turns between planning, testing and implementation. Any useful/recurring tasks, like getting/filtering logs, checking statuses etc get regularly folded into a project cli and agent skills that are local to the project.
I spend time interactively iterating on the plan/design and acceptance tests, then let it do its thing. Usually I have multiple interactive sessions at once that I'm steering this way. 1-2 assigned projects, 1-3 side projects (additional automation, experiments/POCs, workflow/process efficiency etc.). I generally don't do sub-agents. When I need throughput I automate first, frontloading the "deterministic" parts.
Didn't set it up myself so my info will be meagre, but I think it's litellm and a big enough nvidia card. We use it for development but also for internal applications (where smaller models might escalate a task to it when needed).
Obviously not feasible for (just) personal use, the model is too large for consumer hardware.
I subscribe to OpenCode Go when I have a personal projects I need to work on. It's by far the best value since it's only $10 and I doesn't run out of its five hour limit in an hour they way Claude Code/Codex do. Models are more than good enough.
At work, I'm using Codex with the subscription they give me, and I don't really have anything to complain about.
I don't bother picking obscure tools and heavily configuring them since almost all of the intelligence is in the model, and models themselves improve so fast that any complicated setup will be obsolete in a few months. All of the useful behavior just gets trained into the models, and all useful functionality gets added into every harness very quickly.
I'm not using any personal agents since I don't see any use case.
I use a tool I made that lets me orchestrate lots of AI agents (claude/codex/gemini) both on my laptop and on a k8s cluster. I set them up to run on various triggers and access them from my phone/laptop: https://github.com/jonwiggins/optio
1. Claude code, ultra plan, chatgpt go plan (India has a ~3$ plan)
2. Use pay as you yo openrouter with kimi2.6 for an api wrapper interface I have to work with my business APIs, connected to whatsapp.
3. I use opencode alot at my office that connects to a bunch of mixed models gpt-luna, claude. But prefer claude code, because (dunno if others feel it) opencode feels clunkier, and slow. I suspect that's because of the llm routing through my companies infra.
4. Have not made anything work, have tried GasTown, its horrible imo. Also, I think I am not able to trust the setup, and worry about token wastage (which cheaper models do more off when left unsupervised)
5. Open law
- connecting with my personal calender, gmail, whatsapp, side gig infra
- first responder for my business whatsapp for a slice of whitelisted interactions.
2) Not really, the value proposition on the subs are just so much greater.
3) I'm using Claude Code and Codex, crucially thru Muster[0] which I've been working on. It's a herdr inspired Libghostty-powered terminal which understands agents you're running, helps you manage them, supports remote devenvs, and it provides a CLI for agents to spawn, work, and message each other, no matter the harness and across machines.
4) I'm working on my own factory atm, which combines Muster and Kan[1] which is a offline, Kanban-as-plain-files-in-VCS project I'm working on and happens to work really well with agents. I recently added card dependencies to Kan and a kan ready command so I can have director agents orchestrate and hand out tasks based on what's ready and prioritized in Kan.
5) Not at this time. I played around with dots but I haven't been convinced it can do anything for me that my own solutions aren't better at.
Nice! I am guessing the intended use is on developers' own machines, not putting on a server and live co-editing in a team meeting. Pls. let me know if I have misunderstood. Thanks.
Models: I pay Z.ai a lite sub which costs me ~9USD/month and for the last months I've been using `glm-5.3-flash` amost exclusively for my day-to-day. I also make use of free models from OpenRouter for simple CI/CD automated stuff but very little so. When I'm out of sub tokens I fall back to DeepSeek credit. I keep my sub/credit usage public at https://a.l3x.in/usage
Generally speaking, I think I can hardly be happier than what I am already, the quality of open weight models is incredibly good and the only upgrade I could think of is being able to run one on my workstation one day (I don't even try on my Macbook Air M1), which isn't something crazy to dream about, perhaps in 2/3 years running Deepseek v4 Flash or similar model locally won't require too much of an investment in hardware.
It depends very much on the project I'm working on, but recently I almost never run out of tokens before the weekly quota (which seems to be the tightest for me) resets.
I also use some discipline in not letting the context grow too much and usually keep every session very focused, context window fill rarely goes above 10% for me as I prefer to start fresh (perhaps from some markdown artefact saved from a previous session).
If you're used to throw every possible thought to an agent than most probably a light Z.ai subscription is not going to be enough.
I built railvio.com for myself initially, it’s taken me a year to build and it’s a full harness for building applications + fully integrated Jira and Confluence clone. That altogether has become a lovable clone which I’ve used now to deploy a tradingview clone, a risk rating rules engine clone, and an agent terminal that is a stripped down clone of Iterm2 but built with a middle layer so I can connect to long lived tmux sessions. 3x Claude ultra, codex pro, + local Qwen 27b fp8 on Radiance engine on 2xRS700 and ComfyUi on a separate machine with a rtx 4090. It’s been super fun building everything. I’ve averaged around ~40B tokens per month..
$DAY_JOB gives me GitHub Copilot with generous limits so I use GitHub Copilot CLI. We're a Microsoft shop so that dictates some choices too.
My workflow is a windows terminal app with multiple sessions running under wsl2 in each pane since LLMs seem to do better with bash than PowerShell. I don't trust these things much so I've got a careful list of read only tool calls and filesystem sandboxing set up so it can do some useful stuff like read my task board and publish draft PRs without a ton of overbearing permissions prompts.
I usually chat with that for a bit on a task until it's plan sounds good then I have it spin up a Docker Sandbox microvm under Windows so it can safely run yolo mode and make all it's changes. When it's ready I carefully review and iterate on the result until it's ready to publish the PR.
Bit kludgy but strikes a balance of safety and multitasking I'm comfortable with. IME most folks don't think much about isolation so curious to hear if anyone else has a similarly isolated but better workflow. I'd like to try OpenShell instead of docker sbx but it lacks microvm support for windows and I don't want another layer of nested virtualization since my windows machine itself is a cloud VM.
Work also gives me access to GHCP for personal use so I use it for side projects too. Idk what I'd use if I had to pay out of pocket, probably something a cheap Claude plan. For noncoding personal use I mostly just use free tier Gemini. If this stuff ever slows down I'll invest in some GPU nodes for my homelab and run local models.
Our token budget is also high so honestly I don't plan much. Mostly just throw work item and as much related context I'm aware of at whatever is the latest decent model (opus for a while and lately gpt sol) and let it do its thing. If I were more budget constrained I'd plan more up front and delegate bits to cheaper subagents I suppose.
I am using mainly claude code on a Pro subscription.
I also have an OpenCode Go subscription, so I can keep an eye on the capabilities of the Chinese models.
Opus 5.5 medium for the vast majority of coding and code review tasks. I will sometimes save on tokens by dropping to sonnet if I have a well detailed plan built with opus.
I spend a lot of time on the design/architecture of features. These tasks are assisted by Gen AI but I still need to make a lot of corrections. It's my main bottleneck.
For orchestration and sandboxing I've used plain VMs (I have a make VM script which installs Claude in a fresh VM).
I've used Scion at work, which gave me a lot of ideas and got me excited about building a better orchestration, messaging, sandboxing and web UI thingy which I am currently working on (GitHub link in my profile).
My biggest pain point I wanted to address with my tooling is cleanly separating projects and accesses. I do a lot of web dev and sys admin stuff. I'm building Egzo so I can safely grant and monitor access to remote systems and cloud services and keep myself organized.
For my side quests at any given time I am subscribed to one or the other of Claude $200 or Codex $200 or both. I was briefly on 200+500. Getting to be a real expensive hobby.
At my day job I spend about $1000/mo at API prices. So like 5-10% of my hobby usage.
Hobby setup is I have a single "personal assistant" session which spawns sessions for each project or major job on a project and messages with them. I talk to the personal assistant over Signal exclusively.
Inference: combination of Opus 5.5 (best model in the market) and local models (DeepSeek 4.1f 1017 and GLM 5.3f) running on 3x DGX Spark.
Harness: Primarily Hermes Agent across desktop, terminal, and phone (via Telegram). I have also tried OMP but find it a bit clunky to use. Hermes tui + tmux are my favourite tools and Hermes works well with local models.
Orchestration: I haven’t seriously worked with it yet, but starting to use herdr with Hermes. That is as much “software factory” I get.
Claude Pro. Mainly used for developing my apps Pagecord and Feedgrab, but also for personal stuff like diy problem solving. travel planning, data chores, etc.
I feel I get a wild amount of value out of £90/mo (paid for via my apps revenue)
1) 3x Claude Ultra subscriptions and $20 tiers of Codex/Cursor/Gemini, the latter three mostly for multi-provider testing and Google's bundled cloud storage. I basically only use Opus xhigh, but never touch Ultracode.
2) No. The subscriptions are too much value compared to API. If i were to start using any model on API, the first one I'd consider would probably be Inception Mercury 3, for the 1000t/s speed.
3) I use my own custom kanban agent management tool named Kotkit, which I have been developing since March. It has a built-in multi window desktop which I can use from my phone as a PWA.
4) I am currently building orchestration/swarm capabilities into my tool. Currently adding multi-agent deep research and code review workflows orchestrated by my tool, so that they are cross-provider and my agents can trigger these on demand, as well as experimenting having one agent complete an epic ticket on the kanban board by orchestrating task agents (this is all wired up so AskUserQuestion and ExitPlanMode, etc. on the task agents are answered by the main agent, and if the main agent is unsure, it delegates to me). Currently doing it on a >20-task code migration epic with a lot of success.
5) No. AI is not yet good enough at long-term planning for these to be useful for real work. You have to build step-by-step up to it. These agent platforms are good for simple personal tasks, but to use them for actual work would be skipping a lot of steps.
1) i use my brother's claude pro subscription. if i needed to, i would use openrouter + API if i didn't have his subscription, but i don't think i really need it. i've never purchased an ai subscription myself, nor paid my brother for his. i would never use frontier us models for anything serious on openrouter + API
2) yes, when claude starts getting too preachy and the task at hand is important to me i use openrouter + api + chinese models. my computer can't even run qwen 3.8 27b.
3) i use claude code because i think using other ones isn't allowed. i don't really like it, and generally prefer opencode (i've had to use it to do some things for my brother). i'll try using other ones though
4) no, and i didn't actually know they existed. opencode and claude code's agentic things are enough for me, though i don't do actually complex projects
5) no, i won't use them unless i get good local inference going (i.e. never).
2. Experimenting with qwen3-coder-next 80B and a few other local models for coding tasks on my M4 MBP (128GB RAM) - no other models or pay-as-you-go apis
3. I have a custom built setup primarily driven by claude code via telegram with a checked in "brain", a separate "watch" structure of crons that i can setup and tweak, and then a task farm run by that orchestrator with containerized cc (and a mixture of sonnet and local llms based on task) and application devcontainer stacks (if it's coding related)
4. Custom
5. Custom - mainly rebuilding the ideas from there into my own workflows, but not using any of them directly
Trying to split more things out of CC and use it as a pure harness on route to trying other harnesses out like pi, etc -- it's a work in progress as and when I have the bandwidth.
$200/ea plans on Claude and Codex, plus OpenCode Zen for experimenting with other models, plus Qwen3.8-28b locally. (I test-drove DS4 Flash locally in a 128GiB MBP for a while, but the tokens per second were too slow to be useful for more than being a toy.) At my prior anchor client I had an additional max20 plan; at my current anchor client they have a teams plan.
I use both Openrouter and OpenCode to get Kimi, GLM, big Qwen, etc, but not at great scale. I have a security product that consumes API tokens that I built against Opus 4 when that was the strongest available; when I get around to it I’ll likely shift it to prefer GLM or Kimi, slightly for cost but mostly for non-refusal.
I spend a lot of time in Claude Code because it’s the only Anthropic subscription-friendly harness. Additionally use OpenCode and codex. Not much time spent in pi these days. I used Aider before Claude Code’s beta. Claude Code’s plugin framework is pretty decent, so I tend to stand up a repo of custom plugins against every client repo I work with. It’s a nice efficiency layer for me up front and often grows into a bonus client deliverable after a couple months of refinement.
OpenClaw runs persistently, mostly for little calendar tasks and trying to mitigate my ADHD: energy level check-ins, redirects to use attention for things that matter. It orchestrates Hermes under the hood but I don’t interact with the latter directly.
I still mostly just use tmux directly and tell running sessions about each other, sending messages to each other via text files or send-keys/capture-pane. Herdr is installed but mostly unused. The GUI clients aren’t my thing; I’m a born console junkie.
What I really want to change next is pulling this all back into emacs.
For local inference, DS4.1F 1017 quantized to 2.25 bpw from oh-my-pi; this is slow (~500 tps prefill, 30 tps decode) because it's a Radeon R9700 on PCIe4. When I don't need local inference, Opus 5.5 reviewed by GPT-6.1 Sol in VS Code, both $20/month. Mostly I use them to build small analysis tools because I am not really an individual contributor any more. I used to hit both my session and week quotas frequently, but that is much less common since Opus 5.5.
Usage is mainly wiki building around personal info (finances, topics of interest), calorie counting, and doing various things from my dumb phone (rideshare booking, checking MTA Bus schedules, etc.)
1. I use both Melious[1] and Entrim[2]. I have a Melious Pro subscription, €150/month and use Entrim pay-as-you-go API, because it's cheap.
I used to have two Mistral subscriptions, because I needed the extra capacity, but quickly realised I'd need three or more, so switched to the two providers above.
2. Providers covered in the first point. Regarding models I use a combination of GLM5.3, GLM5.3 Flash and Deepseek 4.1 Flash 1017. Which model I use depends on the task. GLM5.3 Flash for general chat and co-ordination tasks. Deepseek 1017 for most coding related work and GML5.3 for adversarial batteries and documentation.
3. Maki[3], although quite customised with various plugins. No python permitted, only Janet for scripting. Extremely limited bash calls, instead preferring to use the built in harness tooling, or plugin tooling I've written. AGENTS and skills used extensively. Mechanical gates also used to prevent certain actions, like remote git pushes without explicit prior approval.
I switched to Maki for the token saving strategies and stayed for the ergonomics, which I think I like, but maybe just feel comfortable with now.
4. None listed. The closest I come is a multipane Zellij setup with different agents running in different panes. Orchestration is still human in the loop though.
For personal use: Claude Pro, Opus 5.5, occasionally API / usage credits. Outside of that, some smaller local models (lately poolside)
I still use emacs, I run agents in buffer or in different terminal tabs. I like pi but only claude is approved at work, so I mostly stick to that for consistency.
I'm use 1-3 agents at a time, not more than 2 on the same branch. I have some pipelined workflows but I don't run swarms.. I'm on budget!
First off, I cannot pretend I have this solved. But I can share one thing that has worked for me, along with codex maintaining it's own local memory (turn it on).
Have your agent maintain a task journal. It is like a more detailed commit history, that includes tribal knowledge gained during the task and additional context, eg.
X is incorrect but we cannot change it due to some other service depending
on it, don't use Y as it is deprecated in favor of Z.
Reset your context frequently instead of only relying on compaction. Use the following prompt to summarize the current context for the new agent:
Describe the current project or larger task in progress, then give a
breakdown of all the planned sub-tasks with status, explain any learnings,
then note what work is remaining. Write this in a file called HANDOFF.md
Give it a try and feel free to share any improvements.
The context should be self-evident from the code alone, and if it is not, it should be in your head and you should write it down again so that you do not forget it yourself.
Being forced to re-explain stuff to the LLM is good, actually, as that way, one actually thinks about the architecture, choices and limitations.
Try to keep context shoved in a conversation thread under a token limit to save tokens.
Ask AI to record in a projects directory memory: LOG.md, HANDOFF.md, DECISIONS.md, a plan folder. Reference code with absolute paths. Keep repeatable processes in skills. Version control it all, and use a dreaming process across multiple sessions to stop repeat mistakes.
That's what's worked for me, but there's many variations of this.
> Try to keep context shoved in a conversation thread under a token limit to save tokens.
For big tasks/features, I'm always creating a new git worktree and I'm using a "status repo" skill that creates a sibling status git repository where everything is recorded.
It works quite well whenever I have to revisit something. It's quite token hungry but I have the full trace of what happened in the past.
Another approach I want to explore is using https://github.com/netdata/netdata (disclaimer, I work at Netdata) to send all OTel logs/traces of LLMs to it. The idea is to replace the status repo skill with something that will allow LLMs to retrieve the history of a given task/feature. It will probably require some "checkpointing" to make this more efficient.
^ This is basically what I was getting at. Thanks for putting it so clearly mate.
> Try to keep context shoved in a conversation thread under a token limit to save tokens.
That is another solid tip. If using subagents, use the main thread for only the conversation + launching subagents. Tell it explicitly not to implement anything, and only support integration efforts.
Doing so keeps the worker subagent's context clean - so it can stay laser focused on the task.
I usually have a “close out” routine that’s something like telling it to record hints to the project’s agents.md. It’s usually pretty good at figuring out what to save, and I usually set up the prompt to point the thing’s attention at the part of the process that made me most frustrated.
1. Claude $20/month and Cursor $20/month (I've been meaning to close it) I mostly use Opus 5.5 on Medium. I've been experimenting with High effort, mainly at work.
2. I have an OpenRouter $50 pay as you go API key that I use for Mercury 3 at this point, primarily for tokens per second. I also use GLM 5.3 when I inevitably get [cyber] blocked (or for reverse engineering in general, apparently) by the morality police at Anthropic.
3. I run things pretty stock. OpenCode with OpenRouter but otherwise I just use Claude code.
4. I've never used these, but I'm curious about them at this point.
5. Not yet, but also interested. My plan is to use Hermes for some personal agent work on a local model, but I have not set it up to test this yet.
1 - Opencode Go $40, Claude $20, $0-100 on Deepseek API
2 - No
3 - I use a lot of Opencode + Deepseek, strategize some of that work with Claude Opus, and some gtm, research, and project stuff in Hermes with its coding agent. When I guide by hand I use Orca Dev, which is lovely.
4 - Claude drops GitHub issues that Opencode picks up on a schedule. It’s simple and resilient enough for me. Hermes has a lot of scheduled tasks to move things forward.. it has gtm and research and dev loops (the latter builds autonomously for future projects vis a vis my attention, I’ve used quite a bit of it).
5 - Hermes Agent, which has become a core tech for me. I’ve tried to use Openclaw but it has not been something I can keep running. It spirals and crashes after a week in ways I’ve never experienced from Hermes, doing the same things. I have little interest in those generalist agents, maybe after a year.
I'm currently using https://www.conductor.build/ for working on multiple projects at the same time with git worktrees and my Claude subscription. The annoying part is that if I have a frontend repository and a backend repository I have to tell each session "talk to this workspace to adapt the backend" if I'm working on the frontend. I wish they had better support for multiple repos in one session.
Separately I'm evaluating https://ampcode.com/ which is very polished, but I feel like the diff-reading part is not very prominent there which I find a bit odd. After watching many of their videos I feel like that's because they are opinionated on this not being necessary any more as the models are good enough. I kind of agree, but still like to give it a last glance before committing.
Claude Code (and Codex too I believe?) can message other sessions these days. I just run one session per repo and they talk to each other directly when the frontend needs something from the backend and vice versa. This has removed most of that back and forth for me and eliminated the need for orchestration tools.
Have a server at home, so mosh into it and run Claude Code at the lowest level. Use Blink off the iPad.
For normal work though, I have a bunch of services set up to run Claude —-remote-control in every one of my projects, so I just use the Claude app on my phone and iPad.
Also maintain and run https://github.com/common-pattern/claudeflow as a service the bigger projects - it’s an orchestrator that reads GitHub issues and starts a Claude one-shot, automatically handling the routine stuff like CI, commits, PR landing flow etc.
For building, I started with Kiro and later switched to Antigravity (Opus 5.5 for planning, Gemini 3.8 for implementation - but now using Argon for everything and quite happy). Claude for design.
Had a Hermes set up with a dual NVIDIA DGX Spark and Deepseek v4 Flash, but then built my own harness (getcinnamon.ai) and pretty much switched to that 100% including family members.
```
nix run nixpkgs#llama-cpp -- server --models-dir models -m models/gemma-4-12B-it-Q8_0.gguf --port 8080
```
open http://localhost:8080.
I also run pi in a docker container to do some vibe agentic coding. But I prefer `cat flake.nix | pbcopy` and paste into the browser. Then I can export the jsonl files and use tools to convert them to md and push to GitHub Pages.
I can do most of AI sessions off-line and only need hotspot to sync with GitHub.
My main use case is to learn and hack nix on Mac. It is fun.
It has a 5 hour usage limit, and if using the wonder model, you basically only get 1 or 2 prompts. However, I find it will basically do unlimited 'turns', so as long as I give a good prompt, I can get my work done.
No ide/harness. I only do light coding and don't have time to actually sit at my computer.
2) Yes, trying different models and harnesses through many providers (free/trial). mimo-v2.6-flash, minimax-m3.1-flash, glm-5.3-flash, deepseek-v4.1-flash, gpt-oss:120b, nemotron-3.5.-super-vl.120b-a12b, and others.
3) Mostly eca.dev on Emacs, and Claude Code at work. Now trying the OpenCode one too.
1, work - claude team pro (not paying for), home, codex pro (for side hustle https://rolloutrhino.com/)
2, no, except at work where I have a whole agentic helpdesk system that self improves, creates documentation, and is quite awesome really. Pretty cheap to run also.
3, my own that I keep iterating on, but I do use the clis, though with codex, the app is just awesome, don't use the tui (its good, but the app is better). Claudes app sucks in comparison, though is often better than the TUI.
4, my own tools, mostly because I started on my own ones ages ago.
5, dots, It's kind of like a project manager for the side hustle, its really coordinating a lot of stuff between all the things I have in flight
Curious what you find better about the codex app vs the claude app? I use both as well, but I find them fairly interchangeable. I like project and session creation a bit more on codex.
1) $100 Anthropic via work
2) codex personally
3) Claude Code + Herdr with a root agent steering everything. Waiting for Rex
4 and 5) absolutely not lol
These comments read like people just bragging about having multiple AI subscriptions but not actually giving any info about what they're using them for
I think a lot of people/companies are using 10x the tokens for 10% improvement. If you're willing to spend a modicum of attention/effort looking at what the bots are doing, one max Claude plan with Opus and occasionally Fable is very effective.
I like using Cline the VS Code extension. It can use local models and local inference with Ollama and tons of external inference providers like AWS Bedrock, Vertex, etc. It asks for permission to read webpages or run commands. You can give it whitelisted permission to just read your files which is what I do. If you want, you can put it into ACT mode and give it permission to write to files but I avoid that. If you use something like Bedrock, it shows you how much the conversation costs. It works great with Claude for coding but if I'm doing something super small I might use a local Qwen or Gemma model.
I use Claude and codex and use my own orchestrator to run them. I also used to have open code for a while but the recent codex and Claude models are token efficient so don’t need additional models.
I use Circus Chief to orchestrate workflows that use various combinations of Claude, Codex, Muse, z.ai, and Grok subscriptions. When one subscription runs out, Circus Chief falls back to the next. You can configure tiers to control how that works.
I've been increasingly interested in language-oriented programming and denotational design. So i get claude to create a domain specific language in Racket modelling the software i'm trying to build, we create nouns and verbs whose meaning fits the behaviour i'm going for, then when we find bugs or notice we need more exceptional (less-elegant) one-off behaviour to tackle edge-cases, we iterate on the language in such a manner that it these are tackled emergently with preferably less concepts, more cohersion and composition. It's somewhat similar to how memory safety is handled at compile time in Rust but for business logic.
Copilot in VS2026 + GPT5.5 configured as a custom model (API Tokens). Same as it was 6 months ago.
I really want to use 6.1 Terra, but the way copilot is set up in visual studio proper, they don't handle reasoning models correctly (or maybe my install is screwed up). Eventually, I will dust off that old winforms project and get my custom harness working with the latest responses API contract.
I prefer (pre) paying for tokens. You can offer me all the proof you want but I'll still never believe that those subsidized tokens are always the same as the ones I get through the API.
Moving towards long-running, asynchronous delegation where I use a locally running instance of Fizzy kanban board to manage and track all work.
When I think of something I might want an agent to do, I create a card. When I want to delegate I assign to agent and move to To Do. I usually try batch things, then use a slash command to invoke a fizzy skill. I do this in a terminal and then close it and do other things.
Couple hours later I'll check the board and see what's in review column for me. Read comments, look over PRs, etc. Agent has broad access to specific project folders and repos.
In general feeling more like working with a colleague than chatting with a bot.
I'm currently on Raven as a sort of "meta harness". The main chat is with the orchestrator, which delegates all work to other harnesses. I just use different seat-based plans. Mostly the latest Sol with my corp chatgpt account.
I get a lot of value from having an LLM to talk to LLMs for me. It can automatically set up adversarial reviews from other ecosystems, or just prompt the implementation LLM to adhere to certain standards. My goal is to have fewer touch points when implementing. I want to speech to text a paragraph or two giving "taste" to the process; everything else is just toil.
I get the same "one-shots" from the vanilla codex as codex under raven. It took some work to get the orchestrator to send good refinement prompts, and it still doesn't have good taste. I need to bring that with my own review of the code. If you notice you send a lot of repetitive refinement prompts it's worth it.
For software development, I am using Kilo Code (https://kilo.ai) extension on VSCodium, with Kilo credits that gives access to hundreds of models at API pricing.
1) Claude Pro (the $20/mo tier). No other subscriptions.
2) I use models on open router occasionally, primarily GLM and Deepseek models. I’ve experimented with Qwen locally but the models that run on my hardware are too dumb for extended use cases and too slow for quick questions.
3) Personal use: Primarily Zed. I use employer paid Claude Code in the day job but it eats tokens at a rate I’m not willing to pay for for personal use. I’ve tried OpenCode but their security track record is uninspiring and I’m way more defensive about my personal hardware than my employer is about their hardware fleet.
1) Work covers my Claude and Codex subscriptions. Each is the ~$200/month tier.
2) I have a "live Zoom call audio transcription" setup that uses Silero VAD, Qwen3-ASR-1.7B, Qwen3-4B-Instruct, and pyannote speaker-diarization-community-1.
3) Claude Code in the Desktop UI using remote control environments.
1) I bought a z.ai 1 year subscription for their mid tier (second iteration of their pricing) for like 500€ total, so 40€/month for mostly GLM-5.3. I suppose flash is there too, but.. eh.
2) I have a 4090 which runs Qwen3.8-27b q4 fine with 180k ctx at q8. The model is capable enough to do things, but dumb enough to not be able to build elaborate nonsense like the bigger models love to do
3) I'm using https://github.com/juggler-ai/juggler which is alright. It could be a bit less vibecoded (and the readme and docs and all would benefit from less LLM writing), but, generally, the opinions shaping it seem decent, and I like just how much control over and insight into what happens it gives you.
I initially tried OpenCode, and, holy hell, what a mess. Do not ever use OpenCode. I also tried goose but it was too agentic for my taste.
4) No. I am doubtful that that kind of workflow can produce software I would deem worthy of production.
I also personally just dread the whole overhead that is tickets, jira, etc, so recreating the same but now you're the manager seems like a personal hell to me.
For 4 years now I’ve tried the latest and greatest once annually before my summer course starts— just to know what to expect and how to instruct. The artificial stuff is definitely getting more refined, but using it doesn’t yet make sense for me personally.
I am committed to learning about it; it would be nuts not to.
Not committed to using it.
Until recently my mental health post-burnout has been so absolutely poor that I have been learning all of this just to try to concentrate on anything so I don’t feel further out of the loop. Most days I still feel like I should quit IT.
In the limited ways I am actually using it in a productive rather than self-educational sense, I tend to be using it to combat burnout: helping get past blank page problems, rough ideation and keeping track of ideas, doing slightly more cohesive research than just ending up with dozens of open tabs and some unclear notes, generating helper processes (like content migrations) that I might otherwise not be able to make myself write, and exploring using it to do the unthrilling work like automated test generation.
I don’t plan on becoming someone who sits back and specifies loops, and that doesn’t fit with the work I am doing. I have far too many disjointed things I have to work on.
But it is definitely a more varied rubber duck, and that has some potential. I feel a little less stuck, sometimes.
Opus 5.5, fast mode, xhigh/max reasoning, ultracode as needed. Usually 5-8 agents going at any given moment. In-house frameworks for Claw-like workflows and agent orchestration.
1) claude, claude again, codex, antigravity, cursor
2) no
3) using a variety of different harnesses managed by codealong
4) built https://github.com/codealong/codealong
5) openclaw not used very much anymore in favor of native harnesses
Claude Code Ultra in Emacs. Split Emacs windows: Bottom is Claude Code, top left is code, top right is db-backed editable spec (propositions and premises) with Clojure like eval that sends spec to Claude Code.
My workflow has simplified a lot over the past year and I personally haven't seen a need for complex orchestration tools or the agent platforms like OpenClaw / Hermes.
Since agents can message each other nowadays, most of the time I just talk to one Opus 5.5 session (because of its 1M context window) and let it talk to all the other sessions, acting as an orchestrator that I use from the cli and the mobile app. I also have some of my agents reachable over IRC through a small CLI I'm building.
I subscribe to both Claude and Codex. Mostly at the top tiers, but I downgrade every once in a while based on how much I'm using them.
I don't use Codex directly as much, even though their harness has gotten a lot better and snappier over this year. Most of the time Claude is the one using Codex on my behalf. I was using this delegation workflow a lot until Anthropic fixed the word salad issues on Opus 5.5. Nowadays I mostly do it to load balance my token usage.
I also use locally hosted models (both via CUDA and MLX), but much less for coding unless it's for asynchronous tasks that I can't use subscriptions for. My main use case for local inference is privacy and tinkering.
I've also been building my own harness focused on security and privacy for the past year or so, but it's still a work in progress.
1) Neither.
2) Yes, I use pay as you go from OpenAI and Anthropic.
3) OpenCode
4) No
5) No
I use OpenCode to update my Home Assistant automations. Home Assistant also does two lightweight checks a day using gpt-5-luna (Look at this photo is the pool cover on or off?). I also use OpenCode for a few other small basic tasks.
This is me personally. At work I have a ChatGPT sub an a, seemingly limitless (I don't use it much tho) OpenAI API key.
two max Claude suscriptions, active dev in vscode to share screenshots, GitHub tickets for memory, basic skills, superpowers,subagentic coding with opus and the main in fable, daily Cron tasks for answering email/linkedin/etc. Desktop Claude for video/music editing . No API usage, expensive for my needs.
Locally (MBP 128Gb) I have MLX-Serve constantly running with Clef (for model routing and decisions), Qwen-35B-A3B and Ornith-1.5-35B.
And I switch to Qwen3.8-27B (the Perplexity Pro version that works well with MLX-Serve) or GLM-5.3-Flash (sushi) when I need to do serious work offline.
As a general harness for coding, etc. I use Swival.dev . Configured to use Clef for model routing, with Qwen-35B-A3B for documentation tasks, Ornith to ask questions about code, and various GPT models from a ChatGPT subscription for more difficult tasks.
For music, pictures, etc. MLX-Serve is all I use, with the recommended models.
It's really nice.
And as a general agent for research, etc. I absolutely love Perplexity. This is by far the AI produce I use the most, ever since it's been available.
It's a little bit sad that they are forcing people to use Perplexity Computer, because it's so expensive. But it works incredibly well.
Perplexity can now run locally, but even though they include the best Qwen3.8 quant in my experience, the Perplexity tool itself has major issues with context management and tool calling. So, when that doesn't work, I fall back to MLX-Serve and Swival with the Chrome MCP server.
For agentic web browsing, I can't live without Comet anymore. So many websites are painful to use, and finding the right buttons to click even to perform very common tasks is a challenge. So, I frequently just ask Comet to do it for me. I also use it to compare products, compare places on AirBNB, write personalized messages on Instagram, etc.
As for an OpenClaw-like tool, I'm using NBClaw. I was initially very excited about it. Look, I can just send messages on Signal and let the agent do work for me! And I occasionally found it useful to remotely fix issues on the host it was running on, or long tasks such as making sure that all my projects worked with a new version of the Zig compiler. It became a replacement for Perplexity because, hey, it's cool, everything's running locally, using my own web browser, etc.
But eventually, I used it less and less. Nowadays, I still enjoy the fact that it's sending me a summary of cool things it found in Hacker News and research papers on topics I'm interesting in. But that's pretty much it. I don't use it much for agentic tasks anymore. It still happens (I recently used it to fix a bug in an opensource project while I was on the bus), but mostly just only when I ran out of Perplexity Computer credits.
I was using VSCode with Claude Code extension, multiple tabs at once, multiple instances per worktree.... then I listened to the Lex Fridman episode with DHH [1] and they mentioned Herdr [2] and I tried it out - NEVER LOOKED BACK! It has been an extremely long time since I completely abandoned and changed my setup - surprised myself.
Herdr is awesome - it pings me when an agent is done or needs my attention and shows me all my terminal panes that are running an agent that I can access with one click.
I am now 90% Herdr terminal, open IDE maybe 1-2 times a day just for more convenient way to look at something. I searched ways to stay in the terminal - found lazygit and love it!
I use long-running worktrees (e.g. named worktree-1, worktree-2, worktree-3, ...) and have parallel work going on in them so I can work on whatever I want.
In each Herdr workspace, I have three tabs: 1st is for ~3 panes, one for Claude and two or more to run the app and basic terminal. Tab 2 for lazy git, and Tab 3 for an extra terminal when I need one.
For example, right now I have ~20 Herdr workspaces open now between a few projects, multiple worktrees per project and a few across remote machines ready in case I need them.
When in the zone, I might have 1-8+ agents working on completely different tasks.
I have a couple remote machines over SSH that I use when I need a task to run and I need to close my laptop.
I have a keyboard shortcut to init my workspace to that setup. lazygit is amazing.
I ask claude to do everything for me, for example: "Let's do it, open a jira ticket, assign to me and put into sprint. Fix it and open a pr"
I create artifacts all the time to share findings with team or as a checklist of manual testing I need to do.
I find that gh copilit does pretty good review on PRs, and I have a claude skill to go address all the copilot feedback (review, validate, reply, resolve, push changes if any were made).
I was on the 200/mo Claude plan but had to switch to the Team plan, so I pay 120/mo for two users and I swap to the 20 plan when I hit my 5hr limit on my 100/mo plan.
My halftop macbook m1 max 64gb is doing all my ai locally, i have not paid for any ai subscription for over a year.
things i run locally using apple mlx models:
general llm,
coding llm,
text-to-image,
text-to-video,
image-to-video,
text-to-music,
everything is holding up well. I may eventually swap the m1 max for a higher memory bandwidth device once they are available for a reasonable price in 6+ years
1) I pay for claude ultra coz the plan rates are significantly better than the api rates, and i feel like i can tokenmax harder than if i am constantly watching my credit balances.
2) yea openrouter free models, usually nemotron coz they have decent availability and I don't get constantly rate limited like with the gemma models on there.
3) honestly still using the default IDE, used to use claude code cli but nowadays i just launch claude remote control and just talk to it through my phone.
4) something custom, i wired it all up in slack so i could easily talk to agents through the same place i do my work
5) Used to use openclaw on a raspberry pi to experiment with like autonomous development but i needed something with a bit more transparency so i moved everything into slack.
1) What AI subscriptions are you personally paying for?
My employer pays for copilot, business plan. It also pays for Gemini.
2) Are you using models outside (or instead) of those subscriptions?
Gemini, company-wide. I mostly use it for general research, making clip art, and occasional programming questions / technical questions. I don't use it everyday, but when I use it, I really like it.
I no longer use ChatGPT or Claude's general-purpose chat bots.
3) What coding harnesses and interfaces are you using? Are you sticking with the default CLIs or IDE integrations, or using alternatives like OpenCode, Pi, Aider, Muse Code, or custom-built harnesses?
IDE integration with Visual Studio, and Copilot code reviews on Github. I mostly use agent mode; sometimes I use ask mode, and sometimes I use plan mode. (One very important thing is to instruct copilot, in .github\copilot-instructions.md, to only use its edit tool and not use Powershell to edit files. Otherwise you need to babysit it and approve cryptic find and replace commands.)
For Copilot in Visual Studio, I almost always use Claude Opus 5.5. Some of the GPT models "appear" to work well, but they leave the code harder to maintain.
We now do a round of Copilot reviews with our pull requests in draft mode before asking others to do the review; and we're not spending as much time "reading" code when looking at pull requests.
We don't (yet) have agents running unattended with high-level instructions and then making pull requests.
4) Are you using any agent orchestration or swarm tools?
No
5) Are you using broader agent platforms like OpenClaw / Hermes Agent / Grok bot / Dots / Muse?
No
------
One important thing: I'm after my boss to dedicate time (1-2 weeks) for me to set up and learn more of the hands-off tools. I love working with co-pilot, but it's best for work where I need to be hands-on. I really need to be able to offload simple work, or work where lower-quality results are okay, to a bot that I don't need to supervise.
I feel like this is a great opportunity to mention https://github.com/regnull/openbot, my attempt to create an open-source tool similar to Grok Bot. You can hook it up to OpenRouter and solve 80% of problems with 1/10th of the cost. I use it all the time. Contributors are always welcome.
1) Right now I have Opencode Go Plus, Mistral Pro, Kimi. It changes. Kimi is not staying and I'm probably getting ChatGPT for a bit, last month I had Claude Pro. I stay around €100 a month total.
I generally use open weight models (Kimi K3.1, Mistral Large 4, GLM-5.3, DeepSeek v4.1 Flash, Mimo v2.6 Pro is the current quiver) but I use closed frontiers enough to stay up to date.
2) I use OpenRouter if I have a specific need, audio transcription (MAI-Transcribe 2) being something I use often enough. Also using it just to test random models.
Not much local anymore except classifiers / decision models that I've played with.
3) Pi all the way.
4) I use Paseo to orchestrate agents, I have a couple plugins I've written for it that manages pipelines and such. I generally start a workflow with a simple skill and if it proves useful I like to make it deterministic if possible. Models be dumb sometimes.
6) No but I have some Paseo plugins for personal use, though it's very limited what I find useful. I also don't find that I want any claw type stuff at all.
1. Personally? Claude monthly and I also keep money in the API service and slowly chip away at it with various apps and toys (I created a Fantasy Premier League app that does research and sentiment analysis as part of its runs and that costs money... but not much) . Professionally I use an unlimited Claude via AWS Bedrock.
2. I use both qwen 3.6 35b-a3b and qwen 3.8 27b on my Macbook Pro (M5 Pro, 48GB, 18/20). I use those with ollama and hermes (and Obsidian, to an extent)
3. Hermes Agent, Claude CLI and Claude in VS Code
4. No
5. Hermes Agent
I use a carefully curated selection of agents for my professional work. Agents to handle adversarial code review, planning, questioning, dev work, etc. I also have custom instructions for various repos that need them (where I can define things like repo norms that are not immediately obvious, data models where needed, etc.
I'm using claude with a pro subscription. Claude Code extension in VSCode devcontainers and Claude CLI in a restricted Windows account in powershell terminal for android development. No orchestration other than a directive in CLAUDE.md to allocate coding tasks to sub-agents.
I've moved all my ssh keys into a key-agent which prompts for key use as I got alarmed by claude attempting to login to remote systems to verify things as it worked through problems.
I have two $100/mo Anthropic accounts (watch out the single $200 doesn't raise weekly limits) and one $100/mo Gemini account. I have Opus generate review prompts for Argon and paste them into agy.
I tried setting up a script so Claude could automatically call agy, but it kept hitting timeouts, and letting Claude control that tends to burn my usage too quickly anyway (and if I'm not copy-pasting into different terminals, where's my value add in all this?).
Argon hallucinates, so I rarely use it for impl, but I've actually had better results with it than Opus for SwiftUI and AppKit work.
I try not to run more than three sessions at once, or I get anxiety about having to keep all the session caches fresh and end up typing silly things into some of them every 30-45 min so they don't expire or auto-compact.
2x Claude Ultra accounts, 2x ChatGPT/Codex Pro. Couple of local hosted models for quick stuff (text formatting, TTS, etc.)
I'm sticking with Default CLI/Desktop apps + built my own to combine them for my specific workflows. Used previously OpenClaw and Hermes but it seems to fall apart when you give it any real work. Ending up giving up because I just don't have to tinker / fine-tune it.
I will say, I enjoy Muse a lot for some reason for personal things. I don't overshare anything I wouldn't mind saying in public.
Within the last month, I basically use Claude Design exclusively for front end work and Codex for backend work. Codex just works but sucks as design. Claude is great at design (with enough references) but can't seem to actually get anything to completion.
I want to make a joke about that last line so badly.
I've had the same experience as you with Claude, and I have to choose a single subscription, so I'm on Codex. I definitely wish I had access to Opus 5.5 right now.
I've got a somewhat sophisticated self made setup:
1a) OpenAI API subscription, integrated into an office suite + media pipeline I wrote
1b) OpenAI codex $100 mo plan
2) Current count, 318 additional generative models across LLMs and then media generation of image, voice, music and video. The LLMs are mostly Qwen variants, and the media models are pretty much all the popular media models. These are organized, managed, and API operated locally through WanGP (open source model host) and two additional apps I wrote. One of those apps is a mobile phone pairing remote live video processor, so one can use their phone like a remote live camera at events, and the other is a plugin to WanGP so my software can securely operate WanGP remotely.
3) I use OpenAI Codex in a terminal as a coding assistant. Through that same account I also run the OpenAI Desktop app, and that integrates with my web apps through WebMCP and two additional types of agents in my apps: 1) all my UIs have knowledge assistants that explain how to use the UI they are associated, and they know more complex workflows that UI is a component. 2) My apps have "method actor agents" that are conversationally programmable experts located on specific interfaces, which users then tune to be experts in their work tasks. That's like a smart assembly line worker waiting to be told what they are making. Then, through WebMCP, external agents will interact with my app resident agents, and together they act. Which is kind of amazing. So, in a way, they create a harness through their collaboration, which also includes the user, collaboratively.
4 & 5) no.
My work interest is in creating an environment that promotes deep intellectual flow. One could use my setup and automate a bunch of things and sit back while their brain rots. Or one can also participate in the work at an interactive creative stream level, where the agents and software are doing work but all the creative and impactful decisions delegate to you. You're in complete control over a swarm of virtual expert personalities that work for you, and you make their decisions. The work is office and media, which is broad.
1. Claude Ultra for personal, Claude Max for work, Codex $100/mo
2. OpenRouter, DeepSeek, Luna, and testing out models here and there
3. Claude Code and Codex for the respective subscriptions. Played with OpenCode and Pi some as well
4. I've tried a number of these tools and settled on my own tool (not released) to oversee my agents
5. Played with OpenClaw and Hermes in the past, cool ideas but execution fell flat. I currently use Dots and Duet for little things (mostly monitoring communications). Dots is much better than Duet today but I hope Duet gets better.
gptel + gptel-agent in emacs, hooked up to a Claude API token and ChatGPT oauth provided by work, plus some custom gptel additions (tools and tool invocation inspection mainly).
At work, we're using Kiro. I've got no complaints.
In my personal life, I'm extensively using Hermes. If you look at some previous comments you can see what I use it for. I genuinely think personal AI agents are the future, the way I feel when I use it is how I felt when I tried out the first iPhone. In 5 years time, I think most people will have a personal AI agent in the same way the most people now have an e-mail address.
1) Claude Code (Claude Teams Premium seat) plus pi.dev (self-hosted GLM-5.3) at work, Claude Pro at home. Opus 5.5 Extra effort for ~99% of the prompts.
2) Not really, messed around with Openrouter and Deepseek for some stuff that Claude refused to do
3) I use Claude Desktop (Linux) and a heavily modified pi.dev at work VSCode as IDE. At home just use the iOS app and Claude Desktop (Windows).
4) No
5) I have Hermes Agent setup at work and home but don't really use it.
1. Claude Max - two (personal and work), ChatGPT plus, OpenAI (PAYG at work), Gemini PAYG (mostly for automated transcription). I could probably cut my personal Claude Max account, but hang onto it for the constant stream of personal project ideations.
2. Yes, per above
3. Mostly Claude Code, occasionally OpenCode and Codex. Rex/Ghostty as the interface to it all.
4. Yes, lightly, a custom orchestrator
5. Not actively, have tried most ex Grok Bot, but haven't taken the time
At work Claude Desktop with Opus 5.5.
At home Deepseek Harness, 5-15USD per month.
I don't like CLI harnesses. coz it looks too monotonous.
I prefer default agent orchestration of the harness, since its their job to take care of it. not something I should spend my time and efforts on. If I did, high chance it will get out-dated soon.
I do not trust agent platforms. Lethal Trifecta comes to mind.
1) None. duck.ai for random questions/research for private use. No personal coding project or anything like that.
2) Yes, Anthropic models usually, at work. Pay as you go API. Company is trialing limits, routers and open weight and even local models. Since the Anthropic 5.5 models none seem reasonable to use any longer (for now at least)
3) Claude Code at work. Tried Aider early on and it failed to even read its initial context (Claude models like a year+ ago). I dropped it dead after that - unusable if you need me to specifically "load" and point our files. That's not agentic coding, sorry. Have not used the actual features my IntelliJ Ultimate subscription provides in quite a long time, though I loathe to tell IT/finance that. I do use it from time to time to double check what Claude tells me. Used `pi` w/ local models (some gemma and qwen that'd run on my hardware. nothing usable w/ the work code base - or much anything for that matter, like I'd not even get through an interview coding challenge with that if you ask me)
Did use Codex maybe like 8 months or so ago to ping pong implementation plans and reviews between GPT and Anthropic models but then just stuck with Anthropic for some reason.
4) Nope.
5) Nope. But we do have our own in-house review agent and software factory agent. Like think CodeRabbit et. al.
I use Orca as an “ADE” (mostly for the multi-worktree interface and mobile app).
In my day job I use Claude Code & Codex subscriptions (Codex is mostly for review, second opinion, and as a backup).
Personally I use both oh-my-pi and Hermes with a Codex subscription. I tend to use OMP more now though, since I find its interface a bit nicer when working on the laptop.
After trying so many different setups, now I just use Claude-- the only thing interesting about my setup though is that I separated a "work-claude" and a "personal-claude" install on my laptop, which isn't officially allowed and took a bit of wrangling. But I basically had to do it legally.
1) Claude code (Max) at work. Claude code (Pro) personal account for side projects.
I do still have a ChatGPT subscription. I'm tempted to cancel it, but I like their app.
Gemini is bundled somehow into my google subscriptions. I use it occasionally, but I'd be just as happy to not be paying for it.
2) I do use Openrouter for side projects that need an embedded AI. For example, on Halloween I made a talking pumpkin to greet trick or treaters as they came to the porch.
3) Mostly code. I've tried muse, grok, gemini, codex, and pi w/ openrouter, but I always come back to code.
4) no
5) I have a grok bot account I use to manage personal stuff. It organizes my email and calendar, keeps me up-to-date on news, does health tracking, trip planning, and shopping research. As an example, there's a website where I order lunch for my kids. I like to know each morning if I need to pack their lunch, so I have grok bot log into the website and sync it to my calendar. I then get a notification in the morning with everything I need to start the day, like whether to pack lunches, what the weather will be, or if I need to help my kids study for a test.
Tried just about everything over the last couple of years. Settled on the unified Cursor, Grok, Grok Bot plan for about $200 / month. Doing everything within herdr for programming and Grok Bot for all non-programming related AI life stuff.
It's a solid groove. Grok Bot helped me get so much stuff off of my plate.
I'm playing around with some ideas on top of pi durable. Cascading contexts exposed over http. Mostly deepseek, glm and kimi in there. https://github.com/resolveworks/key
1) What AI subscriptions are you personally paying for? (Claude, ChatGPT/Codex, Cursor, etc)
No. I moved off of subscriptions earlier this year. I don't trust the subscription model from any frontier lab. Just look at what Anthropic has done. The first time I tripped a safety classifier on a seemingly irrelevant task, I canceled everything.
2) Are you using models outside (or instead) of those subscriptions? For example, pay-as-you-go APIs, OpenRouter, or locally hosted models.
Local first. I have a box running Muse Spark on my desk. My main inference box is a launch day DGX Spark (I saw the hardware shortage coming) and has my entire local AI stack. OpenWebUI and everything that feeds it. I moved away from SearXNG and started using SlopSearX (it's quite good).
Everything else goes to Openrouter. I use few models but GPT-6 Luna is my go-to, sometimes I switch it up. Really depends on the task I need to perform.
3) What coding harnesses and interfaces are you using? Are you sticking with the default CLIs or IDE integrations, or using alternatives like OpenCode, Pi, Aider, Muse Code, or custom-built harnesses?
Github Copilot for work. It's meh. Personal projects and general personal use? I started with OpenCode and have since migrated to Pi (omp specifically).
4) Are you using any agent orchestration or swarm tools?
I thought about it. I haven't really felt like I would derive value from changing my workflow and tools. It took me a while to find a good mix of utility, control, and efficiency and haven't wanted to change it until at least next year.
5) Are you using broader agent platforms like OpenClaw / Hermes Agent / Grok bot / Dots / Muse?
I have a dedicated Mac Mini to run Hermes. It is pretty much integrated into my entire life now. I could live without it, but automating general purpose things is the entire point of why I put that machine together.
How is this different from this time last year? Radically different to say the least. The pace of change has been EXTRAORDINARY to say the least.
1. Claude Ultra + a Codex Plus. Probably going to drop Codex soon, GLM is gradually taking on what I intended it for.
2. GLM-5.3 and OpenCode for cybersecurity
3. Primarily custom built harness. Rust centric, heavy on the production engineering gates and a bunch of graph networks for context control.
4. Custom.
5. No.
My general ethos is that if taste is what is important going forward, my code generation must have it's own taste. If you can download my tooling off the shelf, I am replaceable by anyone who can do the same. My end goal is to drop Claude Code once the harness prices start to drift to API pricing. In the meantime, anything that can be automated into traditional and efficient code will be one more thing in my control regardless of future token prices. When I hit my limits for the week, I spend the remainder of the waiting time going over the framework data to see what are my largest sources of consumption and iterate in automating that more.
I don't even understand how people can spend thousands of dollars a month on LLM tokens since I've accomplished multiple major projects with this setup. In general, Claude's better at frontend and debug and Codex is better at feature implementation and research from what I've found, which is why I have both.
2. No. GPT 6.1 Sol is very token efficient that I pretty much never run out before the end of the week anyways.
3. Default Codex/Claude app. I only use VS/VSCode to hand-edit/git sync things when needed.
4. No. Swarm is a waste of token imo, and a single good human in the loop beats any automated orchestration tool.
1) Claude Ultra, Codex Plus, SuperGrok, Gemini Pro.
2) No.
3) Opus 5.5 on high for all the work. Reviewed by Codex by default. Additionally by Grok and Gemini (agy) for critical work.
I switched from Claude CLI to a custom-built harness because I got overwhelmed by walls of text in parallel sessions. Now I have a canvas per project where I spatially lay out cards. Each card spawns an agent in a worktree. The chat with it is condensed to WhatsApp style chat bubbles with all the thinking and tool calls folded away. The result of a card is a demo, usually a narrated video that I can comment on to trigger another iteration. Everything voice first, I mostly speak. Phone companion so I can keep working on the go.
I used to work with up to 6 Claude CLIs in parallel and couldn't go further. With my custom UX, I can do 20+ with fun. Now the bottleneck is actually my machine, so I added remote worker support to my harness to distribute the load to more machines.
4) No Paperclip or Munder Difflin or anything. I actually think it doesn't make sense to create persistent agent personas. Just spawn an agent when needed, give each full access to all skills and docs, and they'll figure it out.
5) Hermes Agent running, sending me regular news research reports, but frankly, it's just a cronjob and the whole Hermes setup isn't really adding value for me.
1. Claude Ultra, ChatGPT Pro 200, SuperGrok, Kimi Pro, Cursor Pro, Google AI Pro, Muse Code High Usage. It comes out to ~$500/mo, mostly from Claude+ChatGPT
2. I use OpenAI/Anthropic APIs for various workloads (app features, automations, etc.), but not for personal inference/coding. Recently both of these providers started giving $200/mo in API credits, and OpenAI already gives free tokens, so I don't usually pay for this usage.
3. I always use the provider's preferred CLI: Claude Code, Codex CLI, Grok CLI, Kimi CLI, Antigravity CLI, Muse CLI. I use this with herdr to manage everything, and then a small custom app to keep track of my usages since I hit limits quite often.
4. No. Just the CLIs. For very large projects I've found Claude Code to be the best -- it can very easily manage many subagents.
5. I've tried OpenClaw, Hermes, Dots, and Muse. I really just haven't found great use cases for these tools
I use Claude & Codex for most of my personal coding/projects. For the other providers I mostly want to have broad experience with different models/harnesses. It's very interesting to see how each performs and what kind of tasks they do/don't excel out. The only con of having cheap plans is you really can only give them simple tasks.
---
At work I use AWS Bedrock w/ Anthropic + OpenAI models at API pricing. Each employee is given a rolling 28-day budget w/o access to Fable/Astra. I use the same Claude Code/Codex CLI + Herdr.
1) OpenCode Go $10, though I'm just experimenting with it. Before I have just been using the free models available in OpenCode, since sometimes its hard to justify a subscription when you're unemployed.
2) I tried a local Qwen model but my poor old 1660 Ti did not enjoy it, and it wasn't all that useful.
3) I primarily use OpenCode running in a podman container for actually working on stuff. Also use the default free chats with Gemini 3.5 Flash for general searching, ChatGPT for brainstorming, and Claude to answer more code related things if they come up.
4) Nope.
5) Not at the moment. I've been eyeing using Hermes, but since I don't feel comfortable paying much at the moment I haven't really tried.
I'm using Claude Code (Ultra) across a multitude of tmux sessions, over two or three laptops – largely so long running actions don't happen on the actual laptop I take with me (nothing worse than waiting around for an agent to finish before you can leave a café). To make it a little easier to visualise across those sessions I built https://belfry.robgough.net which is like herdr in some ways, but just uses standard tmux.
I also gave in after seeing my co-founder use it, and gave GrokBot a go. I use this for everything that's not coding or design... so marketing, email reading, right now it's currently inputting my meter readings to my electricity company after emailing the building management company and getting photos back. It's super useful at little chores that eat up time. I have it daily check the status of my various projects and services. I love it tbh. I've also built my own version, which I might move to... though now it's all set up it's quite sticky, and it already spotted an error my accountant made on some taxes so it's already saving me money and I'm feeling quite positive about it.
I'm getting a bit frustrated with having to run things on my laptops though -- it feels a little pets not cattle, and wish it was all in the cloud. I'm toying with building out an agent factory, but frankly there's no super-cheap way of doing this so I'm holding off, though I can see me building a tool that connects those laptops to it creating sandboxes with Apple's `container` (which I'm using successfully on my grokbot replacement).
$200 ChatGPT + $200 Claude, I don't pay for it, agents quant trade and pay for their own subscriptions. I then use the rest of the tokens to build stuff.
DeepSeek Flash with kicad + autorouting churned out a couple of working PCBs with minimal prompting, and customized ZMK itself for some custom input devices. At least for casual stuff this is a solved problem
1. Using ChatGPT with a Business Premium seat (not personally paying for it).
2. Also have some API keys for some open weight models in Europe, but currently not using those.
3. I've personally used OpenCode, Claude Code and Codex and of those, I liked Codex the most, till bugs kept on accumulating and they started doing all kinds of shit like that daemon recently. I now switched to https://github.com/capocasa/3code which is a small, simple agent harness but with built-in sandbox, unlike so many other simple harnesses. I really like this one, plus the author is very responsive to feedback
4. No. And I don't think that stuff works in the long run with current quality of code output
I've made a multiplayer environment for me and a few friends, where we can prompt (currently three) emergents and experiment with various models and instructions for them.
I've rarely been so inspired by a project.
It's publicly readable, but writable by a group of ~20.
Codex (Pro x 1), no outside models, custom harness, custom orchestration, custom agent harness.
I avoided building anything custom until I had to, and I avoided adding more than a single agent until I had to.
My preferred way to use AI for coding has been to work closely with a single coding agent with the code open. That changed wth 5.6-Sol and a handful of skill files. I stopped needing to watch closely to keep it from fucking up the codebase. With the right request, that combination will get the job done 9/10 times.
That unlocked a new workflow: as soon as a change is requested, a coding agent reads it and gives me a PR. I want a complete proposed code change in front of my eyeballs by the time I open the change request myself.
So I have a custom solution in place to manage a handful of agents aligned with this workflow.
Custom solutions are so cheap and easy now... it's really hard to put up with solutions that are only mostly right for the job.
I still just pay $100/mo for Claude. Use the CC extension in VSCode. I’m actively investigating various ways of getting a Hermes Agent going on my homelab with either a local model, Openrouter ZDR models, or both but I’m unlikely to pull the trigger on that yet. I would never use Muse, Grok Bot, etc as a personal agent.
I use OC with various subs, and am testing DeepSeek Harness, JetBrains Thinkrail, and Volt. Extending the harness with plugins is my main interest these days.
I mostly work on problems at the intersection of manufacturing, mechatronic design, electronics, software and physics. So I'm not really building software per-se which likely accounts for the probably slightly weird choices here.
1. Claude Max (upsold me as of today); generally Sonnet 5.5 with the occasional Fable 5.1 when it gets stuck or I am executing a high value step; and Perplexity Pro (legacy).
2. Mostly local voice-to-text w/Parakeet Unified EN 0.6B via Handy[0] (awesome; use it!). Not for coding, but for note taking, AI interfacing, and general typing reduction and RSI avoidance.
3. None. I tried Pi and after much kerfuffle I don't like any of the harness interfaces. I find the type of work I do they are not very useful, because I often switch between variant approaches, planning, data, and don't want unrestrained context bloat.
4. No.
5. No.
I would add that I am very interested in obtaining good enough hardware to run this level of model locally at any pace, believe Qwen is basically there but too slow locally to use as a daily driver on my current rig, and can see myself making that investment in single digit years. Not right now, the hardware is overpriced for what it is versus venture funded cloud AI subsidies.
- agents work in their own lxc spawn automatically in a proxmox node, where they can go wild without sandboxing or isolation, network is firewalled too
- all sync sessions and everything to a centralized agentsview in a realtime, where I get to track and watch it from anywhere on the go
- i just ssh into the lxc if i needed anything
- different harnesses depending on the task, maki is great one, but I use others
- I also have other stuff but I won’t mention it here regarding the tokens for obvious reasons
I tried most of the tools like herdr or paseo or desktop clients but they never clicked with me
1. Claude Code with $20 Claude subscription. This used to be a terrible value but now it's a great value, probably because of the IPO. I do a lot of coding and other things and I rarely get close to my weekly limit. I mostly use Opus 5.5 medium. Codex with gpt used to be good value but recently I canceled it cause I kept reaching my limits.
2. When I want a "different pair of eyes" to review things I'm working on with Opus, I use Pi with other model providers and just pay API prices.
3. Deepseek harness with a deepseek API key. Really great value.
4. I have a Muse $5/month subscription that I keep as a backup although I don't really need it.
I also use cloud agents to setup reoccurring jobs. For expensive agent runs, like visit 5000 websites and extract something from each, I would set up something cheap with an agents sdk and for the model use either Muse, Deepseek, or GPT-6-luna using their batch API.
I maintain all my dotfiles and agent files / context in a github repo so that I can switch easily between harnesses and providers.
2) kapitsa + pi + Qwen 3.6 35B a3B q4. pi is the harness, kapitsa is the inference engine
people that I considered intelligent have associated Qwen with "chinese cloud" when it implies not using a cloud at all. I think that's weird. I think there is a "western" centric view that limits perspective to using A cloud, instead of aspiring for locally run
1) I also pay for Claude, I should downgrade again to the $100 plan, thanks for the reminder, but currently the $200 max x5 plan
- then download vmware workstation pro (you ll need to create a free account to get the download link on their website)
- install vmware workstation pro
- turn internet off
- plug your pendrive in with the rufus w11 x64 iso on it (mind ya, this pen drive ll first need full formatting before burning a w11 image to it)
- press f2 to reach bios in vmware settings and make it boot from pendrive
- turn internet off on your machine (DO NOT keep it on)
- install w11 (keep a 100gb for vmware)
- without updates it ll load real fast and you wont need an MS account (rufus takes care of it)
- while it installs, download revo uninstaller pro on a separate machine https://onefiles.net/9d170d01078634b5 (latest version 5.4.x) and put it inside a second pendrive / portable hdd
- once w11 loads, it ll ask you to enter a password, keep existing password blank and enter literally anything for a new password
- voila you are done with w11 setup inside vmware
- now connect the 2nd pendrive with revo and install it (need to clean uninstall many tools, hence we need this one)
- now go to search bar and type check for updates and update everything on w11 inside vmware
- next stop get unigetui from ms app store or from here https://devolutions.net/unigetui/ (it is a package manager responsible for installing just about everything else with auto updation of all packages)
- the open unigetui and search for the following packages vlc, winrar, winhance
- winhance is the fully customizable debloater you ll need to turn off all telemetry (we ll debloat after installing all the other requirements)
- next go to ms app store and install python or open powershell and type python
- the open unigetui and search for the following packages vlc, winrar, winhance
- winhance is the fully customizable debloater you ll need to turn off all telemetry (we ll debloat after installing all the other requirements)
- next go to ms app store and install python or open powershell and type python
- mind you this installs a python version manager that lets you install and uninstall multiple python versions
- after this python version manager is setup, install a specific version of python like
``` py install 3.13 ``` - next we need a node.js version manager, use fnm for it
- install ghidra , ghidra-mcp (mind you open ghidra-mcp first because it only works with specific versions of ghidra and java) this is the reason we installed a java version manager
- run winhance, check recommended settings under optimize and every other tab and restart the vm a few times for full debloat with 0 telemetry services
- congrats you have a full blown agentic environment at this point (if I missed anything i ll come back and update)
Ok I don't understand how people can deal with the information coming from a fleet of agents burning 100s-1000s of €/$ as month.
I have a Claude Code pro subscription, and mostly have Opus as a thinking buddy, we have thoroughly designed a data model together, we have implemented it in Django. Then we designed the RBAC system and now we are making views per role.
It is so intense! Constant decisions, walls of text with pros and cons to decisions. I mean the speed is amazing, the code quality is also amazing, but the amount of information I get to process is just insane, and I enjoy it less than the flow-state writing of code. I think I'll get used to it, and the speed is motivating, but man. I can not picture keeping 10's of agents on tasks. What am I doing wrong? Should I trust the system more? Or am I just mostly working on decisions now and the 1000$ token burners are not bothered with that because they have a full backlog, where I need to fill it myself?
I deal with the overload by making abstractions (a visual of the data model/db schema with color coded relations etc, a table for RBAC, think about permissions in "Axes"... I'm always looking for smart tricks to "think better", to be able to hold more information and reason about it. What must exist in me, what can I outsource?
Btw, as a FOSS person I so hate that I am not owning my means of production anymore, I have a home battery for all my infra, very proud of what I build, but I can't run my agents :( I really hope that will change soon (for privacy and for control reasons.) I pisses me off that I can't get nice hardware, meanwhile I can get a Muse or Dot, which is an LLM with a VM with up to 12 GBs of RAM, for free, paid for by their borrowed money. This whole idea makes me die a bit inside.
For the company I do all IT for we have a Mistral Team subscription, I find it hard to work with anything other than Claude Opus, I really force myself to do it but I often think: Pff, Opus would have gotten this right. But I remember feeling the same when I switched from Google to DDG (3000 years ago) and I never switched back. So it's probably just familiarity.
I also have HermesAgent on a nuc here in front of me, it makes small automations: Scans for positions or projects for my company, parses RSS feeds for relevant cybersecurity news, researches things when I ask, stores notes and articles with I can interrogate as a "body of work" <- love that btw, it's like an external memory where the faintest glimmer of a memory can help you dig up that one nugget of knowledge. Or you can find trends in TFAs you love. You can really get to know yourself a bit better. Most things it makes are LLMs out of the loop, just automations build from simple Unix-style tools that I could have made myself but never took the time for. It's also nice that it solves edge cases itself. It burns less then €3 a month on deepseek or so (just processing public data for me). I keep it on a separate vlan and away from important stuff.
I don't think you are doing anything wrong; it is overwhelming. But I also don't think "trusting" the system more would help. IMO, now coding is a bit more like maintaining critical infrastructure; it may be built quickly, but you need to keep an eye out for random decisions agents are taking that will cost you (money, security...) in the future. I have seen Claude making many sneaky decisions and (maybe) coincidentally burning more tokens by doing so
You use Claude as a "thinking buddy", meanwhile those who spend $500+ on LLMs have shut their brains off, and have delegated all thinking to the LLM, producing tons of slop.
herdr, opus 5.5 (max), codex (max), and I am still refining how they all communicate, currently codex became unwakable at some point while claude can be woken with a message just fine.
moved to openclaw for everything, with occasional claude usage for tricky bits.
I have $200 of both codex and claude subs.
Openclaw is the answer to question "do X if Y happens". An agent per each type of Y channel manages the flow and memory. And user input is just another case of 'Y'. It is a very powerful model I am having lots of fun with it.
the plumbing is a chore though, thinking of trying out grokbots to see if they solved it.
1.a) At work: I can choose between Cursor and Claude on a meter. Been using Cursor for a long time and happy with it until they started to force Grok down our throats. I was happy with Composer 2. Recently switched to Claude Code. The company also runs some open weight models in-house and I have been using them through Claude Code itself, with VSCodium as my Code Viewer. I almost never need anything stronger than Sonnet at work - 80% of which is YAML and Terraform these days (or rather markdown, if you really think about it).
1.b) For personal stuff I maintain an Opencode GO subscription, USD40 tier. I also have a Google One subscription which is supposed to have some AI capabilities, haven't tested it yet.
2) Besides my Opencode GO subscription, I sometimes add a few bucks to use Zen. It's rare: my USD40 is usually fine for the month. I've recently started using Nvivia NIM free tier models for some time insensitive tasks too. I have a Qwen model available on my Mac (M1 PRO 64GB) but haven't found it to be useful, as I need my laptop for everything else I'm doing beyond inference.
GLM 5.3 Flash and Deepseek 4.1 Flash do most of the work for my personal stuff. Nemotron does the time insensitive stuff, which is mostly summarizing text. I've been using GPT Luna 6 a lot because it's currently cheap as hell, so I'll often talk to GPT Luna and hand off coding to GLM or Deepseek. I find that Luna communicates MUCH better than the Chinese models - I assume because it does not go through the extra language barrier step between English and Chinese. GLM, Qwen and Deepseek are incredible tool callers, but their prose is insufferable.
3.a) At work I'm using Claude Code for the Anthropic models and Opencode for the self-hosted ones. Ditched Cursor recently. I've been meaning to try to integrate the self-hosted models into Claude Code but haven't yet, so hand-offs between the two harnesses involve a lot of markdown.
I've switched out of Cursor because A. they've been bullying us into using grok for everything and B. The company setup would not support the use case of mixing the self-hosted and closed models in the same harness - this limitation is on our side.
3.b) For my personal stuff I've been using mostly Hermes. Some actual coding goes into Opencode, often piloted by Hermes itself. I'm using Zed as my code editor, but with the AI features disabled. My opencode subscription is actually configured in Zed, but I prefer the UI without them.
4) No. I don't really need them. I run 5-6 parallel sessions at most, otherwise I can't keep track of what's being done anyway. Each parallel session is on a different task, so I generally don't need them to spit enormous amounts of code. I rarely have to wait more than a few minutes for a task to finish, and I'm always thinking or writing or managing something else in between anyway - not looking at the screen waiting.
5) Hermes. Generally happy with it - despite it being extremely unstable between updates, with frequent regressions, weird UI bugs and half assed features (bot mode, anyone?). I have 3 distinct instances (laptop, home server, VPS) doing different things - all hooked up to the same memory provider (honcho) under a single user profile, and a handful AI profiles on each. All have limited access to my Obsidian Vault, which is synced between them using Obsidian Sync.
I generally use the laptop instance for day-to-day use: planning work, assigning tasks to myself and to the other agents, capturing ideas, light coding, etc. The homeserver instance manages a limited set of features on the server itself, from upgrades, documentation and more devopsy stuff to organizing files, managing transcoding queues, etc. The cloud hosted version automates is not very used yet. I plan to have it pilot an n8n instance I keep to automate some boring stuff, like filing weekly tax records and automatically forwarding bureaucratic e-mails, nothing fancy.
> 5) Hermes. Generally happy with it - despite it being extremely unstable between updates, with frequent regressions, weird UI bugs and half assed features (bot mode, anyone?). I have 3 distinct instances (laptop, home server, VPS) doing different things - all hooked up to the same memory provider (honcho) under a single user profile, and a handful AI profiles on each. All have limited access to my Obsidian Vault, which is synced between them using Obsidian Sync.
I am just getting into setting up Hermes on a VPS. If you have any high ROI tips/tricks/advice on how to make this work well, please let me know!
Embarrassingly, I don't have one yet. I'm mostly still writing code one keystroke at a time. I've used Github Copilot to translate between languages and look things up, but haven't trusted it to write new code or fix anything.
Claude Fable is now good enough that I should be using it. The latest Claude version is said to be comparable to a mid-career graphics programmer, which is about what I need. Six months ago it wasn't good enough to use for such work, and probably would have made a hash of my code. There's been progress.
But Anthropic's business practices are scary. They have a reputation for cancelling accounts for random reasons, offer rather bad terms, and have no customer support. I have a strong aversion to services with a reputation for cancelling services arbitrarily. They will apparently cancel based on their system's guesses as to what you're working on. Sometimes they guess wrong. What I'm doing (metaverse clients) is so far out of the mainstream they may guess wrong.
I'm currently stalled signing up for Claude because their verification email system can't get to my inbox. (Some problem with their using Constant Contact, a spam system, to send transactional emails to my own domain, which, long ago, I had blocked at the Constant Contact end. I think. The verification email never touches the inbound mail server.) It's very important to Anthropic that you sign up with the "right" email address, and it may not be possible to change it, so using some disposable address is not a good idea.
Suggestions?
Just use a random gmail address, it works.
Rolled my own for a Claude Max sub, basically tmux and an elaborate MCP for persistent semantic memory, Forgejo, etc. (https://github.com/ASIXicle/Coterie)
I sense more people will do so as you can fairly easily tailor any harness and MCP/LXC combo to suit the way you work. It's a bespoke future.
Subscriptions: Personal: Claude Pro 20$/mo Work: we have access to some Anthropic and OpenAI models and open-weight models, like GLM
Outside: I have some open weight locally on my 64G ram M5 macbook, mostly for personal stuff that I don't want to send to cloud (or claude). Mostly using Gemma through oMLX, because it feels more street-smart than the more coding focused models, experimenting with Qwen-flash-next through Strata, because the streaming of experts really pushes what size of model I can run
Harness: Claude code, mostly, I trust anthropic, and like the remote integration with their mobile app. Being able to say "prototype me this and that ui" for personal project and seeing it and testing it on my phone is awesome.
At work for the api-provided llms, I am trialing opencode and pi harness, but don't have much to say, we mostly used claude-code up til now.
For local I mostly use Hermes, not for any of the big features, just because I want a nice chat interface (I am mostly talking through my health problems and trying to get my appointments in order) that already has basic tooling integrations out of the box. I.e. getting OpenWebUI have search tool was pain. Adding i.e. mcp for my todolist or calendar to Hermes was smooth.
It used to be more interesting and elaborate, switching between different models. I used to use Cursor, then Codex, but since Opus 5.5 came out, I just use the Claude IDE. I got rid of all my skills and just set it to medium. Then I vaguely talk to it about very high-end things that I want, and it comes back with magic. Rinse and repeat.
I'd say the biggest trick is just to tell it what complete looks like. Be very specific about what you want it to come back with once it's done. That gives it a good idea what the goal is. It'll automatically create subagents and all that stuff. Although if you have money to burn and you're doing something large, you can tell it to just go wide or boil the ocean.
Also, make sure you're working on at least five different things. Otherwise, you're sitting around for one agent.
I can do 5 things concurrently for about a day max. 3 things concurrently all week and my weekends are spent untangling my brain. 3 things per day on Mondays and Tuesdays and 2 things per day concurrently on Wed-Fri and I feel ok but crave holidays.
I probably have some kind of pysccosis. I've had a personal $200 plan from one of the AI providers since March 2026. I'm never not making things now, I haven't rebooted my laptop for weeks because there is always an agent building something.
I cycle between 4-5 work projects and 2-3 personal projects. Every night I tell me agents I'll be going to bed and to not ask questions and keep going. I've always loved building things, but this feels more like an addiction than healthy sometimes.
They're not all done, and half the projects I bench, but this is where my personal projects live. https://wooten-labs.com
love the website, but quality control is important when working with these agents.
ex. i tried cool walk (https://wooten-labs.com/coolwalk) and the 'Download on the App Store' does nothing at all.
respect. this is how i've been working the last 4 months or so. it's a bit crazy making and i do think i'm enjoying things less because i'm preoccupied, but i sure do feel productive. lol.
> I probably have some kind of pysccosis
No you don’t. You are doing everything right :)
Nice website :)
Honestly it does sound like an addiction. Touch some grass...
I had the same issue. What you need in a management system for your claude sessions.
I firmly believe that those using out-of-the-box Claude Code or Codex are likely getting the most cutting edge features of agents anyways.
Yep, Claude & Codex desktop are innovating extremely rapidly. Some of the features like code review are hidden under commands, but basically obsolete whole heaps of AI code review startups when you use them.
I just ask Claude code to do a code review of PR X and it spawns a sub-agent. No hidden commands or anything special.
If I hint it might be security sensitive code, it bumps the effort level and burns my usage.
But it works quite well.
The findings are often debatable or unreachable states but it does find stuff and improves the overall quality I find.
What's funny though is that it almost always inevitably finds something wrong, even if you just ran it in another window and fixed everything it found. There is really a point of diminishing returns.
I’m a user of both and don’t really see the need to go anywhere else. It all just sounds like mumbo jumbo.
As a solopreneur of a boring SMB.
I still do coding and build my own tools but I don’t see the need for agent swarms or Muse or Grok Bot or any of that.
The reason why I stay away is because I like to do QA on the slop these things generate. Having them run amok just makes me nervous.
Same journey for me, although I still use skills. Claude Code is fine; the remote-control feature is why I stick with it now. Generally anything in the cloud doesn't have a proper GPU that supports WebGPU, and as nice as SwiftShader is often you want to actually run things on a real machine. remote-control is by far the easiest way to do that from my phone.
> Claude IDE
Claude Desktop's UI seems pretty similar to Cursor now, almost. I'm still on Cursor but if I use it daily I often hit usage limits before the end of the month. Would you say using Claude directly is better bang for buck?
What do you mean by Claude IDE?
Claude Code Desktop. https://code.claude.com/docs/en/desktop
Yeah, I used to hate the Claude desktop app's Code tab, but it's actually pretty good nowadays. I actually prefer it now over claude CLI.
I use Claude Code Web almost exclusively for personal projects -- I’m guessing that’s pretty much the same interface but in the cloud -- and hardly ever run anything locally. Nicer UI than a terminal, real sandboxing, and it doesn’t tie up my computer. It’s just a remote agent that can make GitHub PRs.
Yeah I was thrown off by that as well. I assume they mean just Claude TUI..
I have a similar story as OP, for me its the Claude Desktop, not claude code.
INTERNAL: Local Agent on PC via PI Coding agent OS / kernel Arch Linux, 7.2.11-zen1 CPU AMD Ryzen 5 5500, 6 cores / 12 threads GPU 2 × RTX 3060 12 GB (24 GB total), driver 615.78.08 RAM 62 GB DDR4, 3 GB swap Model Flash-Next (IQ3_XXS, 75.8 GB) Context: 65,536 Prefill - long prompt: 732 t/s Tokens per second output: 53 Uses: Advanced embedded design, testing and coding
EXTERNAL: Deepseek-V4-Flash on Opencode via Deepseek.com with prepaid Tokens Usage last 30 days: $4.37 USD Uses: Application building for Flash-Next above.
1. None 2. see above 3. Opencode and PI 4. no 5. no
Claude + Codex, both from work.
Just those.
https://github.com/oscscull/quarterdeck built my own harness. I had trouble with both claude and codex. In particular, managing multiple tasks in a kanban, syncing between my machines, making sure work is done on the right machine and can be triggered from either. Making sure I ping reviewers, meeting reminders, todos. Parking tickets til conditions are met or time elapses. Automated notifications when I have meetings. I have just kept adding features as I want them. This kind of software is so easy to make nowadays, I don't really think there's any sense in "launching" it for others, but I'm using it all the time. I highly recommend doing so (building your own personal work management software in 2026).
I guess this is also in some sense an agent orchestration tool, I don't use any of those.
Don't use.
> building your own personal work management software in 2026
Agree. I made my own "project tracker", does exactly what I want. Zero clutter. Zero mention of deadlines or due dates.
Personally: I have a XTX7900 card which can run some decent models, eg, Qwen 27B and Gemma. For bigger stuff I have the Ollama Cloud subscription to access (mostly) GLM. I have a server with OpenWebUI, and for harnesses I use OpenCode, but I have been experimenting with pi.dev to create a workflow that takes the ideas from the Rational Unified Process (RUP) so that everything gets properly documented and reviewed by a human before getting to the implementation phase.
At work: Claude Code subscription, mostly with Opus.
I'm juggling multiple claude ultra subs, multiple codex $500 a month subs, antigravity for argon, plus API rate deepseek 4.1 flash, mercury 3, and abliterated large v2 for red teaming.
I work across multiple models and harnesses using a really simple redis-backed message bus called "nova-bus": https://github.com/mas-bandwidth/nova-tools (open source).
With just this message bus, I can coordinate AIs to work together at a normal pace (worktrees, subtasks etc...), but past a certain point the AIs can't keep up and I need to use a sprint card system I developed to coordinate work.
With this sprint system I'm able to work 200 cards wide (each card is around 30-60 minutes of LLM work in a session) across a fleet of 20 machines linked together with tailscale and ssh.
Full system is here: https://github.com/mas-bandwidth/nova-sprint (open source)
It's been very effective at getting high throughput, but afterwards I find that the resulting code can be a bit of a pastiche from different models working in isolation, thus a good pass over working directly with models to clean up and harden the code is necessary.
I find that using nova-bus to coordinate multiple model/harness review of codebase in very effective at increasing code quality. Each model sees things the others don't.
Is this amalgamation working on a single project? What harnesses do you use? How are you using mercury 3? How much and what kind of work are you sending to deepseek?
Can't work on a single project, Amdahl's law gets you then. Try to work on as many different unrelated projects as possible, so I can get maximum throughput.
I break work up into four loose tiers:
1. flash - deepseek flash 4.1 / luna / haiku
2. pro - mercury 3 / terra / sonnet
3. heavy - opus 5.5 / sol
4. frontier - fable 5.1 / astra / argon
The decision of how to classify each piece of work is done by the coordinator (typically Opus 5.5...). If a card fails at a particular tier, it's going to get escalated up tiers until it gets sorted.
Each card has multiple readers (code review). I find mercury is very good at reading quickly. Deepseek as well. I will be honest, I think for 99% of stuff I need to do Opus 5.5 @ medium is ideal. I just use the additional models for different points of view, especially the absolute frontier models fable, astra, argon at max effort for security audits and code review.
The sort of work I pass through this system, it's mostly library work for multiplayer games.
For example, my work on schema, my protobufs/flatbuffers language replacement for multiplayer games, parallelizes very nicely across 15 languages on the sprint system. https://github.com/mas-bandwidth/schema.
See roadmap here: https://github.com/mas-bandwidth/schema/blob/main/ROADMAP.md
Very difficult to coordinate this amount of work in || without some sort of structured work system.
Even more AI coding work here if you are curious: https://patreon.com/masbandwidth
Thank you. Moo looks really cool.
I did some prototype games, one thing that worked really well for me was having a deterministic replay system where the agents could drive the gamestate via recorded inputs. Kept me from having to do any playtesting or debugging in the main agentic loop.
You'll like this too then
https://github.com/mas-bandwidth/fixed3d -- port of box3d to fixed point :)
https://github.com/mas-bandwidth/fixed -- fixed point math library
Cool! I’d forgotten until now, but I had that exact idea back in the stone age, 10 or 15 years ago, and started working on it but got bogged down. It was clearly feasible but a huge amount of work, too much to justify the cost. So it’s a perfect job for coding agents now. Nice one!
I do you think now need to watch out for the opposite problem. These things become so easy to build that the temptation is just to go ahead and build. But unless you really do need them, all that new code just goes to waste. The cost to produce has become very low, but sometimes the value can end up being negative.
Live view of sprint system as I work tonight: http://37.27.135.206
Here is a summary of the comments from this thread. Updates will follow to include latest comments too.
https://techaddo.com/posts/ai-agentic-setups
At work we have both Anthropic and Copilot subscriptions, plus Google AI as it’s included in our business plan anyway. I have personally been using the Claude Code Desktop App and was happy with it until they made it harder to pick a branch for each worktree! Which is why I tried JetBrains experimental agent IDE, ThinkRail, and I love it! Super easy to start worktrees from any branch and the spec driven development experience they promote is great!
It also let me connect to Requesty for my hobby projects and with that I can easily try any model, from any provider. That gave me a much better appreciation of how models perform, how fast and how costly, thanks to Requesty’s super awesome dashboard which is a pleasure to use.
I was only using Opus 5.5 before as I wanted the highest quality, but using smaller models you can get much faster work, and the difference in price is dramatic. I have now settled on Haiku 5 for simpler things, it’s nearly free while being really smart still, and for big things I use Opus 5.5 still but also Google Argon 4, it’s incredible as well and it seems to write more human text than Opus.
We also have a framework that lets us launch Claude (we could use any harness but for now only Claude) in our Kubernetes cluster within a devcontainer. That lets us ship heavy work off to another machine so our laptop is not overloaded with agents. It works great but we are not open sourcing this framework yet so I cannot link for now. There are some alternatives out there but most of them are out to make big bucks instead of providing a small but powerful solution for this problem.
I build my own coding agent : Spurcode https://github.com/getspur/SpurCode which enable to support me to orchestrate difference other coding agent : like Codex + claude/Goose ...etc
Also : give out product agent notebook workbench : https://getspur.dev/spurlab to make all coding agent need to explore/grouding before coding with notebook
I'm using a cli I wrote to orchestrate per-project VMs for sandboxing using libvirt with COW shared golden images to minimise the disk overhead. The VMs get "rebased" periodically onto new golden images for the rootfs, but keep a persistent home FS for longer. I'm going to clean-up the documentation and open source it at some point soon.
I was using oh-my-pi, but currently playing around with Hermes as an alternative.
Exclusively using open router at the moment, mostly on principle around ZDR, and freedom to experiment with harnesses - but financially I'll probably have to compromise on this soon. GLM, Kimi and DeepSeek are the main models I'm using, with some OpenAI usage.
Tmux for remote access and session durability. I'd prefer to find a setup where I could access the same session through a web interface and seamlessly switch back to TUI, ideally spawning sessions through a web hook or similar. I'm going to take a look at Hedr to see if this can improve this.
Probably the two major things I'm planning to do next are looking at adding a bifrost VM that will hold all API keys, and centralise management of MCP configuration, potentially use it's skill repo stuff as well.
I also want to start controlling egress more - probably iron proxy, though tbh I'd like a solution that was relatively liberal whilst still mitigating exfiltration and prompt injection risks. Maybe I can find a curated allow list to use as a base.
Finally, I make a lot of use of open-webui, but it's disjoint from the agentic setup currently - it'd be nice to find a better way to share information/research from it into the agentic sessions.
1) Codex $200 * 2, Claude Code $200, Muse Code $50, OpenCode $40. I run them in 6 machines (2 Linux, 2 MacOS, 2 Windows)
2) No. I try my best to use oauth
3) Claude Code + Custom Harness(set of skills) + Mod. This orchestrator (I call her as "Team Leader") orchestrates other machines with Tailscale. - I've used Claude Code (Opus 5.5) by model satisfaction and mod.
4) None
5) Dots and Instinct. Main use case is research, code search and raise issue, PR reviews.
I use AI primarily through work, I don't pay for a personal AI sub.
I do have some OpenRouter credits but I mostly use them sparingly to try open weight models out for small personal projects.
I currently interact primarily with Claude Code, but most of my async workers are on Codex.
The orchestration part is custom. I talk to a "chief of staff" Opus instance to ingest various workstreams (Jira tickets, alerts, personal ideas, etc) and turn them into beads. It launches workers to tackle individual beads using Remuda, an agent launcher I wrote at work in 2025 and open sourced earlier this year: https://github.com/yendo-eng/remuda. I configure profiles in Remuda that launch containerized agents using my preferred harness and model. I update these profiles as the pareto frontier shifts. Currently it's GPT 6 Luna max for most tasks.
It monitors the workers and has them go through at least one round of agentic code review. It surfaces PRs for me to read once they've passed agent review. It then either ships them or hands them off to my coworkers for codebases where I'm not allowed to just merge PRs.
Around this workflow I had it write a small collection of skills and scripts to grease the wheels and make repetitive actions more reliable.
Since Remuda supports several terminal multiplexers, I've swapped out the backend 3 times: tmux -> herdr-tui -> herdr-core. There's experimental support for Rex but I'm waiting for it to mature a bit before I try and daily drive it.
It's not nearly as polished as some of the software factory setups and fancy GUI orchestrators I'm seeing, but it's light, hackable, and so far the bitter lesson has kept this working really well for me so I haven't felt much of a need to try anything else.
I have been trying out Muse for a while but it hasn't found a solid place in my life yet. Still too unreliable and too many privacy concerns to justify giving it a lot of access.
I have had a $20 codex subscription. I initially just used codex cloud but more recently have been using opencode with the open chamber ui over the top of it.
Codex usage limits have really gone down hill so I also now have a $10 opencode go subscription on which I pretty much just use DeepSeek 4.1 flash. I’ve found DS to be a great model - probably somewhere just below GPT 6.1 Sol but at a fraction of the price.
In terms of workflow, I discuss features with chat gpt (because it’s free and it has access to my project’s repo, read only view of prod logs etc, so can make informed decisions). Once we’ve agreed on how a feature should look I ask it to create a spike github issue. I then get another agent in opencode (usually DeepSeek flash or Luna) to break it up into implementation tickets.
When I’m ready to build the feature I can give open chamber a list of tickets to work through using the task plugin, so each ticket gets a fresh context. I try to build directly on the main branch and try to ensure we can always deploy to prod even if a feature is half finished (using feature flags etc to switch it off until it’s ready).
Once a feature has been implemented I get GPT Sol 6.1 to do exploratory / regression testing with a real browser against a test environment. This usually teases out a lot of edge cases I hadn’t considered and also gives me a chance to improve usability etc.
It often takes a few iterations of this (exploratory testing, fixing issues that are found, exploratory testing again…) before a feature is ready to release.
I can’t believe how far things have come since August 2025 when I first tried out codex.
I built a Qt GUI-based harness, Humbug, that runs on MacOS, Linux, and Windows (I have reasons to use all 3). Has only 3 runtime dependencies other than the Python standard library: https://github.com/m6r-ai/humbug
Usually have 3 or 4 instances running at a time and lets me use LLMs from 9 different providers (and not using their libraries either). The low dependency approach was slightly limiting early on but now makes it very easy to extend and makes it very easy for an LLM to integrate new features or optimize existing ones.
I typically use Claude Sonnet for document-related tasks, and currently DeepSeek 4.1 Flash is my current go to for coding (it's insanely cheap on ollama.com) and I find it's still fast and reliable at very large context lengths. Previously I used GLM 5.2 a lot and before that was Sonnet for most things.
Humbug lets one AI delegate tasks to another and it can choose different models/providers when it does, so works nicely for getting second opinions on things. I haven't yet found a need for an agent swarm though - more often the parallelism comes from working on multiple projects at the same time.
One of the biggest unlocks for me in the last few months is having AIs write ADRs to capture design thinking as a project progresses. It lets new AI sessions discover the intent behind the current work and surface any potential conflicts that might arise with new requests.
One fun feature is the AI can control pretty-much any UI surface the user can control too, and it's quite fun to watch it edit files (you can watch it scrolling around the file in an editor) or use a terminal. It's a nice visual way to see what's going on.
One of the biggest unlocks for me in the last few months is having AIs write ADRs to capture design thinking as a project progresses.
I find this works great as well. I try to do it on personal projects (without AI) with mixed success. Because it turns out I'm human. But when I started doing this on AI projects it made a huge difference. Features that I'd proposed and not implemented were picked up by the model and cross-cutting considerations were raised. Just insane how that raised overall output quality.
For me the answer is twofold.
Regarding commercial solutions:
- Subscription: Google AI Ultra (testing for 1 month rn to see what i can get out of argon, after that probably going back to Pro package)
- Harness: Antigravity IDE : from what i have tested on commercial harnesses this one i liked the most so far
Regarding selfhosted solutions:
- Inference host: Im owning a 5090 and im running the models locally - mostly use qwen 3.8 27b q5 atm, its a bit "slower" even on a 5090 but it fits the vram i have available while also being quite good in terms of result quality
- Harness: I tested alot and ended up writing my own harness ( https://github.com/laughingmandev/loa ) since none of the ones i tested (yes i tested openhands, hermes, etc...) did work well for me. While my own harness still needs some enhancements, its especially made for long running / complex tasks while beeing hands off (not permanently prompt jokey'ing ) and sturdy enaugh to work well with selfhosted/local run inference.
I use the gemini4argon/google agent atm to work on a c++ simulation and image asset generation for a game im working on. The complexity of this stuff just is to big for a local qwen 3.8 to fully grasp (its not that good at c++ either). Local image diffusion also is not that great.
For other stuff of low to medium complexity im working with my own agent/harness/setup - especially if the data is anyhow security relevant since im not leaking anything outside. From writing scripts/tools to analyzing infrastructure or discussing architecture.
1. claude pro, ollama pro and opencode go 2. Openrouter for testing 3. I use VScode and claude code 4. I've actually built my own for agentic workflows - https://vectorstep.io 5. I use openclaw for my personal stuff, VectorStep for proper trusted workflows (and production workflows)
1) 2x Claude Max 20x + Chatgpt Pro 2) I dabble with open-code and hermes especially to try out the free models. 3) I use vanilla claude with an orchestrator on top 4) https://github.com/mhalstrom/maximizepm 5) I use Hermes. - I use hermes as a "free" agent platforms as you can easily connect your chagpt sub and claude max plans come with api usage you can hook up. In my website https://app.toolscaled.com/docs/agents/how-agents-work/ instead of re-implementing agents and paying API costs you can hook up your automations to push to hermes, claude routines or chatgpt's equlvalent essentially getting agents for free which you can then incrementally convert them to workflows.
- Antigravity (AGY - Google Ultra $100/month), mostly gemeni flash but I benchmark Gemini Astra sometimes. My main driver because it’s pretty good at search compared to setting something up with DSH.
- DeepSeekHarness (DSH) Desktop using:
—- with pay as you go DeepSeek tokens, more expensive to use during my night/Chinese daytime, otherwise very cheap
—- oMLX on my MacBook driving mostly a Qwen 3.6 35b A3b MOE (mostly used for benchmarking a less capable model)
I mostly do custom subagent programming and can trigger subagents in AGY and DSH easily enough. DSH supports a Python coordinator orchestrating DSH subagents as conversations via RPC. AGY only supports a sub-agent orchestrating other sub-agents but it works well enough at this point. I plan on releasing my subagent framework soon (called autofab in pypi), which is basically a specification driven code development agent system (using existing agent harnesses rather than implementing my own).
I tried goose, OpenClaw, Claude, codex, etc…and they didn’t really work for what I needed. I might get a $100/month ChatGPT subscription to go with my Google one, primarily to get a different perspective and benchmark openAI models on codex (Claude doesn’t really support what I need it to, otherwise would go for that instead).
I use pi as a harness, and Paseo as a Mac UI (because I need a Ui).
I tried a bunch of models, but so far deepseek flash from October is just dirt cheap; and roughly on par with the frontier open weight models.
It’s still somewhat behind closed weight models (I only know Gemini Argon) in that it produces lots of text that is approximately correct —- frontier models seem to be very good at producing concise text that’s very correct.
I pay per token, using an EU hosting provider (melious.ai).
But I can fill in the gaps myself, so for me developing for fun, this is enough.
Are you happy with melious.ai? Are there any other EU inference providers you have tried?
I was on tensorx.ai first, but melious’ deepseek offering is way cheaper. I switched to them, no issues so far.
They are also backing Ecosia BTW ;-)
I use Melious as well, haven't had any issues so far.
Nice! Just heard about them here for the first time, but will give it a try. Model selection looks good and pricing seems fair!
Currently I am using the regular Claude extension with VS Code, but one of my friends is working on a new harness called NodeCode that doesn't require regular compaction, so I'll be trying it out as soon as I can get some time to do it: https://nodecode.ai/
DeepSeek Harness using DeepSeek Flash API
Nothing else, DS is smart enough to do everything I ask for and more, never hallucinates and never disappoints. Apps in Rust, Swift, Node, we even designed a new language better suited for LLMs, a compiler, and it runs beautifully. Haven't coded a single line since May 2026 and I will remember that date as the day AI took over my mundane job as programmer and promoted me to architect or director of my own apps
A custom plugin to ring a bell whenever a task is finished and a gong when a question needs an answer, that way I can play backgammon or mahjong in the meantime
I can't be happier
DSH user here as well, what new language is better suited for LLMs? Is it open source? Is it still as readable as rust or python?
I’ve found Pi with DeepSeek Flash is excellent, and seems to be about 90% cheaper than Opus or Fable.
+1 for DSH, it really should be getting more attention here. The only problem I have with it is getting search to work, but now DeepSeek has a solution, though you have to pay for it soon.
1. Claude (max) + Opencode Zen. 2. Tried some other models via OpenRouter but the pricing just doesn't justify usage sometimes. 3. Claude + OpenCode harness. Using Paseo sometimes to coordinate parallel tasks although I quickly run out of memory or grind to a halt doing this. Mentally I can't coordinate more than 3 or so coding tasks at once, and physically my Macbook pro suffers when going more than this. Convenient. 4. Forget swarms, I can't even handle 4 parallel tasks. 5. Tried, and these seem deeply overcomplicated, very AI-built, and ultimately not useful.
I figure my experience might be uncommon here so I'll offer it: None, no, none, no, no.
There are perhaps more dozens of us than we expect, eh?
This thread is super biased by the power users, they are the most excited to share
What is your job and how do you do it?
I built my experimental system and I'm using it. I'm genuinely very happy with it. If I had more compute and could run a real, full 30B model locally, I'd be absolutely thrilled.
https://github.com/8lamaster8/GIN
1) I use 2x ChatGPT Pro 200, and a Claude Max 200.
2) I use other models via Open Router, but ChatGPT beats them in value (at the moment) thanks to Tibo's resets.
3) I mostly make Claude coordinate a group of Codex agents. They've set up a communication channel between them, and Claude acts as the coordinator and a second verifier. This saves me time because it's easier to talk to one agent than a group of agents. Codex parent session does only "comms" with Claude, and all the other agents run as sub agents.
4) For non coding usecases, I use https://www.ctlflow.com. This lets me run dozens of agents on tasks like equity analysis. The information a swarm of agents can surface from a large data store can be quite surprising.
5) I have tested OpenClaw, but now I just use ctlflow.com or claude + codex.
> 4) For non coding usecases, I use https://www.ctlflow.com. This lets me run dozens of agents on tasks like equity analysis. The information a swarm of agents can surface from a large data store can be quite surprising.
Can you tell me more about #4 regarding what your heuristics, methods, and use cases are for equity analysis?
I ask because currently I am building a narrow use case Python app using Claude specifically to filter equities that may have an asymmetric potential to buy LEAP options exceeding 1-2+ years.
I am not a finance/quant bro, but I do want to get better at leveraging AI to be a better investor.
Thanks!
> 3) I mostly make Claude coordinate a group of Codex agents.
Same here. Are you just using a skill or something else. Beyond direct CLI calls to each other, I’m using contextify to share past sessions between them.
I'm personally primarily using Claude Opus 5.5 via Claude Code on a Pro subscription. In my timezone I seem to get an incredible amount of usage equivalent to about 2 full working days per week, letting it do all the coding but focusing it frequently. It blows me away how good it is and how much usage they are giving us.
At work I'm afforded a Claude subscription that is 5x that much usage and I simply cannot use even half. I'm not coding at all anymore at work, just driving Claude. It only needs a bit of direction.
I also use OpenRouter with OpenCode, and a homegrown harness I wrote initially without AI help but later on with AI help. But I haven't used this setup for a while. I swapped models based on the churn as reported by artificialanalysis.ai but they don't release as much data via their free API anymore.
I also can run Qwen 27B A3B on my NVIDIA 5060 Ti 16GB but I don't bother with that anymore either.
I don't use orchestration or swarm. Occasionally Claude spawns a few sub-agents as it needs more data and is preserving it's context window.
1) Claude Max 5x, ChatGPT Plus (for minimal usage of Codex, mostly fired as a Pre/Post PR hook for Claude Code to get reviews from GPT models locally
2) None for now.
3) Claude Code CLI, but not always in a terminal I'm sitting in. I talk to it on Telegram and Slack for different projects.
4) Something custom that grew into a project: Spool (https://github.com/enes-alatas/spool). It runs long-lived Claude Code loops, each with a name and a mission. Each one wakes on its own timer or when someone messages it, works, and sleeps, and the session resumes with its context. Each loop gets its own container with an egress allowlist, and they message each other and me in a Telegram or Slack group. A web control room shows what every loop is doing and what it costs.
5) No, Spool covers it for me.
I use free models in opencode, mainly DeepSeek, but it turns out my laptop can run Qwen3.8 flash next at 60 tok/s via Strata. So I mostly use Qwen now in a customized version of pi. But when I need the GPU to be free, I have to use free models in opencode. I don't use any orchestration - free models don't have high enough limits for that, and local model has only 1 conversation cache slot, more won't fit to my RAM
What are your laptop specs?
I have a CLI based hybrid setup on my MBP (M5 48GB) and it has proven to be very effective.
1. I have a $20 Codex subscription that is more than enough for my work. I use a 75-35 model where I used higher models (Sol/Astra for 35% of the heavier tasks and lower models 75% for other tasks. Sometimes I also use Free OpenRouter models for minor tasks.
2. I use many local models like Qwen3.8-27b/Gemma-4-26b and some 9B models. I also heavily use voice/vision models like Kokoro, QwenTTS for marketing and product videos.
3. Pi exclusively, which I have tweaked for my hybrid setup.
4. I have my own custom agent orchestration that routes between Codex and local models based on the task.
5. I used OpenClaw for a bespoke market research task once, it worked flawlessly.
I have both platform.deepseek.com (hooray off-peak pricing) and the $100/m Claude subscription and boy do I find myself mostly using DS-flash pretty much every time.
There's something to be said for the speed with which it gets word done - I take fewer "it's compiling" breaks, and lose my train of thought less frequently. What's crazy is that the speedy work it does is pretty much always correct - which is a reason why I don't feel the need to spin up claude-code and "ask the expert".
This speed & good-work combo is the DS signature IMO
For my harness, Deepseek Harness + DS-4-1-flash combo is almost too good to be true
Claude 100$/mo for personal projects.
Opus 5.5 for 99.9% of tasks. Fable for the rest.
Harness: Claude Code plugin for VS Code. At some point I guess I'll switch to just the claude code app but tbh this is super ergonomic for me.
No agent orchestration besides literally just letting claude spin up agents for me.
No broader systems.
And I'm _massively_ productive with the setup I describe above. I say this not to gloat, I'm sure others have great methods but I do say this in case anyone is feeling overwhelmed by the options out there. My 2 cents is don't overthink it.
I'm a complete newbie on this: how do you make Claude spin up other agents? And what do all those agents do separately?
For example, say you want to search for bugs in 100 code files. Ask Claude to “use subagents to look for bugs across these files” and it will likely spin up 5 agents, each taking a different section of the codebase. Theoretically in the limit case this is 5x faster than Claude doing it without subagents.
Any kind of problem shaped like map reduce can be accelerated this way: Looking through your banking records, searching through a database, asking Claude to find that one weird file you remember from three years ago but forgot where you saved it, etc.
Just know that there’s no quicker way to hit your limits than to use subagents. So use them if it’s paid for by your work or whenever you need the 5x speed boost.
Thanks for the explanation. I'll give it a try whenever I need to do something similar.
You can tell it to do so, but what you'll find is it knows to do it when it needs to. This is another case of "don't overthink it". Just let your agent decide when spinning up sub-agents makes sense.
If you focus too much on optimization, and you do so in the wrong ways, you'll actually find you saved no time/money/tokens vs just letting it decide when to do it. This applies as much to Agentic engineering as it did with programming pre-AI.
tl;dr it will literally do it for you, and it'll tell you when it does so
You tell it to spin up subagents, they do whatever the parent told them to do, usually a segment of a task you gave it.
"use subagents as applicable" (and maybe "use sonnet for simpler tasks")
Thanks. The Sonnet part is smart to save tokens.
1) none
2) Local LLMs on an M1 Max (Unsloth, LM Studio, occasionally direct llama.cpp). Muse Glimmer (because it is good at what it is optimised for: interactive coding sessions). Qwen 3.8 27B occasionally and a couple of other things for fun (an LFM MoE, experimenting with Nemotron Bolt 74B-9B, enjoying the really super Gemma 4 12B QAT when working on creative photography research or non-code planning).
No openrouter (yet) but that will definitely happen.
3) Currently focusing my attentions on Pi via Paseo (and trying Pendant). Pi is just much better in smaller context sizes, and it is the right fit if you are keen to learn from first principles.
I tried OpenCode first. I couldn’t quite get on with it and what ultimately made me switch was trying to understand its terribly messy source code.
4, 5) I am mostly using AI as a coding sidekick so I don’t really have much need for orchestration yet. I did some experiments with Osaurus (really quite impressive) and I actually think that is almost enough by itself, except that I want a little more control over the sandbox container.
I feel that my usage pattern is still settling, because I am not keen to surrender all of my coding to the machine. But I am trying more purely “directed” coding to learn the process, and finally beginning to get a feel for how I will work with that.
t3code on my mac and omarchy, connected to my VMs with tailscale, that runs inside my own proxmox cluster.
I get to separate my work, have a lot of compute, manage it from anywhere, and also have my agents go bananas in their VMs.
I also run 1x Claude Ultra, 1x Claude Max, 1x Grok with X Premium, and 1x OpenAI sub.
I think the biggest leverage move for me was to get the VMs. Now I get to group my work better, and give my agents a lot more power by running more freely.
t3code has been a solid boost as well. It did hurt really bad tho, because I have been a terminal nerd for a long time. But the speed I've been able to achieve across a lot of projects has been crazy with t3code.
What about t3code makes it better than herdr or something else? Is it the cross agent messaging ability?
thb for me it's been the workflow around it. Things like: - Pulling up my phone, dictating a quick thing and sending it of to work on one of my servers. - Shooting things off to claude code, codex, grok, so fast. So well integrated. - Hitting snooze on a session for next week - hiding sessions in snoozed and working, so I can focus on what needs me now. - Dragging filed straight into the chat - Quickly looking at the diff on the side and go through files with the "viewed" button as in github. - ctrl+j to get a terminal quickly to do something fast - ... So many more but this is top of mind.
I just get so much more done, faster, with less context fatigue than before because of how it's laid out.
This is the video that really flipped me. https://www.youtube.com/watch?v=dLhcLqoff6k
Was a hard hit to my identity... But the gains was just too big.
My biggest improvement lately was a cloud vm. I start it when I start working and shut it down when nothing is running which makes it fairly cheap.
On there i run claude code and codex via ssh and zed‘s ssh connection feature to take a look at the code where necessary.
I map a range of ports so I can take a look at previews.
Agents just continue running when I close the lid.
Cursor + Grok for serious engineering work.
Local Qwen and custom harness to experiment.
1. Claude Ultra. I've never had multiple subscriptions from the same provider, though that has looked more attractive lately as I
2. Yes: I often have personal tasks that Claude refuses to do, so I have to use Deepseek-v4.1-flash or GLM-5.3. I typically stay with OpenRouter, but I've found a discount provider that offers a 2/3 discount and have migrated to them. The catch is the data is probably sold, but given that it has saved me $300 this past month for non-sensitive work (game decompilation), the tradeoff is worth it.
3. Claude Code for Claude. I used to use OpenCode for my API usage, but there are too many issues which have gone unfixed and too many weird issues that I became tired of working around, so I switched to Oh my Pi. I'm not a fan of it either, so before my next big run with open weights models, I will probably rewire Claude Code to use these Chinese models.
4. Claude Code's subagent feature and workflow (subagents, but queuing up >200, and running 8 at a time) are sufficient for my needs. I had Claude build some bespoke workflow orchestration tool for managing many OpenCode agent instances, but it is a single shell or Python script that is essentially the equivalent of Claude Code Workflow.
5. I don't have a use case for them, so no.
1) 3x $100 , swapping between openai/anthropic
2) DeepSeek when even OpenAI complains about cyber security stuff.
3) Pi in Emacs (Pillish) and a yet-to-be-published pi-deck on top which makes it more like herdr with more features.
4) custom pi orchestration. Mostly 2 tools: `do` (fork conv with question) and `delegate` (new pi session with question)
5) no
I have a $1k/month subscription from cursor for work. depends on the task priority, complexity I select between opus 5.5 and auto (grok). I'm an escalations engineer
pay-as-you go on openrouter. use latest DeepSeek or Gemini.
I created my own minimalist/task based harness that checks out a jira issue, downloads attachments, runs sandbox docker containers with configurable gh/local repo as volume mounts depending on jira labels, can load agent instructions from local/remote depending on jira type (triage-bug, dev task, reproduce-bug, security) has scoped/temporary access to targets (e.g. dev k8s cluster, aws/gcp/azure account, ssh, splunk). dynamic secrets loaded from external secret management system/env vars. since a lot of the corp repos are littered with skills, rules and other crap that bloated the context, I also set enabled a way to ignore repo/org agent instructions.
I have 2 repos for storing my personal/corp agent instructions and I 2 separate harness configurations.
ChatGPT Pro 200 x2
No
Codex
Codex
Dots
DSH + 3xSpark + GLM5.3 Flash
I dont think i would need anything anymore. Its doing alot more than what i did 2 years ago in within 2 months.
My current setup is Herdr + oh my pi + pi. Then I mostly use oh my pi's "prewalk" feature with Opus/Sonnet as planner and Deepseek Flash 4.1 as executor. I use OpenRouter to access the models.
This setup is working really well for me and provides a good balance of cost and efficiency.
Just pi with some custom mcp for search and web. Oh and Hermes but limited use
Currently on a GLM sub and will need to decide whether to renew in 3 months.
People are complaining about the new plans usage limits but I do have a -50% discount so idk
Gemini Pro with web chat interface. I upload files, ask for changes, copy & paste the files back, and test and vet the code changes.
No other models or software.
I use a custom Daemon. I tried OpenClaw and Letta and it was like emailing your AI. I have a 64gb Mac m1 Max and the largest model I could run in openclaw was a 27b because its not built for local AI, it's built for API. I wanted something that you could chat to, collaborate with, build, evalutate the code, make corrections, build, repeat. I couldn't find anything that didn't come with massive latency so I just made my own. Now I'm running a 125B A3B model at 30 tok/s using SSD Streaming and my agent is instant with no latency and can do everything I ask. I use Aider for programming, exa for searching.. I'm happy with it.
I'm locally hosting variants of Qwen and Gemma 4. I've also tried inference services such as baseten and novita, but wasn't satisfied with inconsistency in performance and pricing.
For software dev, I primarily use my own harness/orchestrator (that I shared recently), and sometimes Pi. I switched away from Cline + VS Code primarily because code is being generated so fast that I needed to optimize for review.
My harness is modeled similar to my day job: dev team works on stories, tracked on a kanban, and I'm a code-review gate. For example I scrutinize control surfaces, but instead of reviewing the implementation: I skim it and ask questions.
Every now and then I check to see if other harnesses have what I'm looking for, but none seem to have a combination of:
- Strong sandbox - Particular UX (for mobile, code-review, basic IDE, and task management) - Code checkpoints - Great compatibility with local models
1) Kagi Ultimate. I don't use any subscriptions for agentic coding, just search. I evaluated using Perplexity Pro for a while, paying for both subscriptions at the same time for a few months. I find it hard to describe the experience of using Perplexity Pro beyond making me feel physically ill when I used it.
2) I primarily use Qwen 3.8 flash q2 on an R9700 with 4x256k context. I also have some Openrouter credits as an escape hatch if I need it (usually to work on my local AI configuration, sometimes to speed up certain tasks).
3) Pi[1] with a suite of extensions and context files that I have built out myself for the most part, with a few external extensions. One of the main external tools that I really like for coding is openspec[2]. I switched to Pi from Codex because I liked the minimalist philosophy behind it, and because at the time I was using a 7900 XTX (which is now my primary desktop gpu, and I also use for auxiliary services like whisper and decision models), and I was having a hard time getting 100k context with any reasonably useful model. The default prompts of Codex ate into that a lot, and Pi was a much better match for the resources I had.
4) I use Herdr (tui)[3] to manage Pi agents. All agent to agent interactions that I do are ad hoc. I'll do things like tell an agent to spin up work in several other Herdr panes that I have opened with Pi in them, or I will do things like tell an agent to inspect the work of another agent and make sure that it continues to stay on task.
5) No.
[1] https://pi.dev/
[2] https://openspec.dev
[2] https://herdr.dev/tui/
I pay for Claude Pro ($20/mo)—using Claude Code with Opus 5.5, and Google AI Premium ($20/mo) for Gemini Pro with Gemini 4.
When I write code with one, I use the other for peer review.
1. I've been using the $100 github copilot tier but am looking for something cheaper, as my usage is waning somewhat. I like to hop between models for comparison, but openrouter always seems to end up being way more expensive for the flagship models. I also keep a GPT Plus subscription for voice, until someone has a comparable offering
2. No outside models yet. Working on a setup for that. Sometimes I throw some money at openrouter to check out the mega-cheap options.
3. At home I really like goose (TUI), but have been messing with opencode (TUI mode)lately which is also nice. At work I use Zed with a corporate claude subscription and it's a good setup.
4) Not really. I tried some of the opencode swarm plugins awhile back but they didn't seem to work right. In general I don't really have the energy for messsing with a new plugin ecosystem. I also like to work in a _tight_ loop, conversationally as things are being built, steering all the decisions. I have next to no desire to hand a larger task to a "team of AIs" to replace my input with more AI.
5) No
I drive a lot and spend a lot of time talking to GPT voice workshopping ideas, talking at a conceptual level, looking up things I'm curious about, and discussing design and problems I'm having, sometimes even work problems. I use goose for all kinds of stuff, mostly infrastructure and debugging for now but I've written a couple programs with it as well.
I do not like terminal tui much as I think it's a poor fit for a coding agent.
But I do use claude code tui because I have a sub: it is a good model and I like its aesthetics. If I could use the sub in my own harness I would. It's also a foil to GPT's models that overengineer everything.
I have a Codex subscription that I use in my own harness, dreamcoder.ai, which has a tiling window manager and all the other mod cons you'd expect a frontier harness to have.
In addition to that, I run Qwen 3.8 27B NVFP4 with ninfer on my 5090. It's a great model for local development. I also use that in dreamcoder.
> I do not like terminal tui much as I think it's a poor fit for a coding agent
I recently moved away from using tui to directly interact with claude code.
I still start a session in the terminal but then ask to start a simple jira- like ticketing system.
When I have a request I start a new ticket and claude decides to either do it or launch a sub agent to do it.
When claude is done or has questions- he replies in the ticket. When claude has a side-findings he creates a new ticket for me.
That's a really interesting workflow.
I also do orchestrator like work that way. The key difference is I'll have a master chat and let it delegate work as needed to sub agents.
I have skills, MCP etc for things that need it, and custom-designed sub agents for a couple of very specific, niche tasks. The rest use generic delegator, explorer, worker sub agents that just take normal instructions.
Can you not? I built a custom harness that uses my Claude subscription. You can pass -p I think is the flag, to feed it a prompt. And it’s been a little while so I don’t fully remember how but you can also have it print the output to stdout. I even pass it custom tools as args and it works with the cache, and you can even append messages to conversations
You can work around it like that for sure; you can call it from the CLI, and you get free credits also to spend.
But I meant more like, you know: hey just use the whole sub in another harness. That you can't do.
1) Opencode GO 10euro (claude max at work payed by empl) 2) Online opencode 3) Opencode harness is better at keeping your costs low but claude code on ultra code is simillar to agent orchestrator 4) Agent orchestrator 5) Ive tried hermes agent but my little server cant hold it :(
Claude Ultra, No, Claude Code, Claude Code can orchestrate agents, No, but I run Claude Code on a VPS of its own for some things.
We have an OpenSWE agent on Slack at work but it seems horribly expensive and slow for anything I want to do.
1. Claude Max with Opus 5.5 and the ChatGPT $20 plan, I have a few thousand bucks of credits since they reduced the Codex plan so I’m using it for all the bulk data tasks that it would be a waste to use Claude for.
2. ChatGPT Image still looks really bad to me so I use Gemini Image. I tried Argon but it really didn’t work for me so I just use the image model at API rates.
3. Just the default Claude/ChatGPT desktop apps. I used the command line tools for a while but the desktop ones are much easier to use when working with images, and you get the usage down the bottom.
4. No, I think this is all a big gimmick. I barely use subagents. It makes sense to have multiple people at a company with different roles due to their different capabilities and experiences - but when the models are identical clones then why not just have one do the whole job? I only use subagents for bulk data tasks because otherwise they get distracted chitchatting with each other.
5. No, I haven’t tried them yet. What’s the benefit over ChatGPT or Claude?
If you use claude code on ultracode it already spawns subagents by default now, i think agent orchestrator has the only benefit of ofloading some tokens and making the subagents to not get distracted by the bigger picture but i cant see any use case of the broader orchestration on a project that is already solid and you just push features and debug, a single session give you more ownership of what is build
Yeah, the ability of the sessions to message each other turns it into a mess. They all start messaging each other with unrelated tasks and you end up not knowing what each thread is meant to be working on.
I notice that single-threaded agents get stuck with things sometimes like writing some tool they need inefficiently so that it takes 3 hours when it could be 3 minutes (meaning that a task ends up taking days instead of hours, unless I intervene and tell it to make a fast tool instead) but subagents just compound this as they burn through tokens without realising the problem.
How about, what sandboxing are you using to prevent the LLM from doing something you don’t want? I’m using nono.sh and have restricted network and file access. I use nono’s proxy so the LLM never gets access to api keys. I create a separate api key for each service specifically for LLM usage, and I restrict those keys as much as possible
I built something which auto-containerizes the agents for you (gvisor isolation, with credentials injected via proxy and url allowlist), but still allows orchestration between the agents, and a desktop app as well: https://github.com/bsklaroff/yaac
I wrote my own bash wrapper around bubblewrap and pasta: https://github.com/andrei-dubovik/jail
The wrapper also keeps history, memory, etc. split by project folder, so agents working in different projects don't know each others' history. If anything, `--resume` is a split second easier.
I have been working on a shared folder that sync between agents (https://www.filestash.app/agents/). The control plane [1] handles pretty much all storages protocols / platform (S3, SFTP, NFS< SMB, ....), you add the controls / permission boundaries and everything is auditable + traces, and the sync client is at the edge [2] to sync the data it actually has access.
[1] https://github.com/mickael-kerjean/filestash
[2] https://github.com/mickael-kerjean/fdrive
Tried nono.sh but it seemed cumbersome. I think it lets the whole agent run in the sandbox, so whenever you need to change something, you have to restart the whole thing, if I remember right.
Also tried Docker Sandboxes which worked fine during the try out but my trust in Docker products with regards to security is low in general, so I'd wanted to wait a bit whether problems show up.
Then used Codex for the built-in sandbox but it's buggy like hell, documented features just do not work properly. Nobody there every seemed to actually try using the stuff they are documenting as working.
Ended up with 3code right now, which also has built-in configurable sandboxing.
try https://www.droneship.run
1) Home - Claude $20, Gemini web, ChatGPT web. Work - Claude $200.
2) No.
3) Claude Desktop (home/work). Cut and paste when using Gemini or ChatGPT
4) No
5) No.
Claude Code Desktop on Windows: I like it. Simple UI that I started using about 2-3 months ago and have grown with it. Multiple sessions, virtual groups for sessions, etc. I can see my manual workload of managing things is getting larger so I want better project tooling. Some of the default selections tend to stick with what you selected first and I dont like those aspects (new sessions seems to default to your original/first session folder - so we need to clean that up or create some built-in /new-session skills, etc)
However, what I really want now is voice integration without having to click on the speaker icon and talk to it (or a way to keep the speaker icon depressed) . I want the ability to get off my chair and away from my screen and talk to it. This part is easy, I think, and would be a good start. To close the loop I would like to hear the feedback. However... getting the responses to have a 'conversation modality' instead of a 'verbose text modality/Claude speak output' may be a bigger challenge. The workflow will never be 100% display free. But if we can get the audio aspects worked out, projection could be to a pair of smart glasses (XReal, Apple, Meta, Google) - activated only when needed (read something or see a UI design, etc) and then pushed aside.
Anyone one have a minimal keyboard/display approach working for them?
1) Only ChatGPT Plus at the moment cause I'm between jobs and couldn't justify all the subs. Grok gave me three months at $30 when I tried to cancel so there's that too.
I wanted to really keep Claude but ChatGPT voice mode is too good tbh and no other comes close. I wish I could use Opus 5.5 for coding.
2) OpenRouter PAYG, mostly for experimenting with Chinese models like Deepseek and GLM, usually for stuff that Codex or Grok refuses to do.
Had been looking into NEAR and Venice for completely private information processing (like emails) but haven't set them up yet.
3) Codex, and OpenCode as harnesses basically.
4) I really wish I could decide on some agent orchestrator without trying them out for weeks. I like my tools highly powerful and trying to find recommendation for this online is hopeless because it's full of opinion from people who never used a terminal before.
5) Grok Bot for running experiments with supplements, calory tracking etc.. Tried to set up Hermes a couple of times but hated it every time. Just gave Rakazo a try yesterday and still hating it.
Gosh, I dream about the time when you could look at a github repo and make a decent guess about the quality of the software.
I use Paseo for mobile / desktop UI, and as an agent runtime for some automation in a dedicated Hetzner box.
I try to keep the harness setup simple and portable, I use Claude Code or Codex with their subscriptions, I invest a lot of time in my skills which are shared between them.
Disclaimer: I am the Paseo maintainer.
Deepseek/Deepseek Harness with about 30 plugins and a lot of customization. I tried IDE integrations with Jetbrains for a few months and hated it. Now I enjoy juggling 3-4 projects at once. Costs maybe $20/week. A lot of jobs you can work at for a long time with little cost because there's a 99% cache hit rate.
Sometimes I despatch it on open-ended tasks that I'd never take on otherwise; for example I have a little fleet of 10 agents digging up medieval Japanese Buddhist parables, translating them from scratch, and feeding them into a custom interface optimized for language study.
Most of the time I give it a moderately detailed description of the scope and key/unique features I want in a new project, and it banks on work from previous projects to know what sort of deliverables I want, track and review its own mistakes etc., then I steer it from a rough prototype into something more polished. I find it equally useful for coding/application development and research/synthesis tasks.
1) Subscriptions - None Using OpenCode BigPickle 2) Models - OpenCode BigPickle and deepseek v1 3) Harness - OpenCode as primary , but started liking Pi recently because of all the customizations 4) None 5) Played around with OpenClaw for a really long time , now been exploring the Hermes Agent
I use T3 Code as my front-end for all my coding agents on all my personal computers and my VPS.
For the actual agents I use Claude Code for my Claude subscription, Codex for my OpenAI subscription (I usually only keep one of those at a time), and pi for my OpenCode Go subscription.
I mostly use Sonnet 5.5 medium and Sol 6.1 medium for everything. But lately I have been using DeepSeek 4.1 Flash a lot and find it very capable.
At work I use T3 Code as well, but it isn't connected to my personal machines or subscriptions and use Github Copilot in T3 Code via ACP. Model choice is similar, but for more involved tasks for architecture analysis or design, big features that touch a lot of interconnected subsystems, etc. I'll turn up reasoning effort to high or use Opus 5.5 on medium or high. But when actually implementing, I try to break up my tasks into smaller pieces and the models usually do well even without reasoning turned up.
Most of my work goes into rewriting prose in reports, commit messages, PR summaries, etc. because all models still struggle with that. (Claude models less than GPT ones in my experience, but still ...)
Where do you find that T3 Code is strongest and where it is weakest?
I just recently started using T3 Code and it's strongest points are having a really good workflow, interconnecting all agent harnesses so you can easily orchestrate across them and they can easily communicate, and truly phenomenal remote sessions. All while having a singular UI/UX experience which I actually really appreciate.
It is probably weakest at customizability and making sure you have parity in all your agent harnesses with skills/etc, granted I just set an agent to ensure that so the point is a bit moot.
Overall I have really been enjoying the inbox style workflow and it has been a joy to work with. It jives with how I like to work across streams.
It has no weak points in my experience (compared to other harnesses). It is so good that I just can't use anything else anymore without feeling frustrated.
Theo of T3 Code has made a recent video about its features.
A few examples of why I like it:
- Not having to use multiple agents (you can switch them mid-conversation with automatic handoff if you want)
- The way it handles environments (I can start a new session on any environment (machine) with two clicks (new thread -> pick project on any environment)
- I can ask for visual responses (uses a T3 Code tool and renders a html page inside chat with cards, charts, etc. that follow the currently selected color theme (even if you change the theme after the response has been generated)
- MCP tools like browser, webfetch, etc. so I don't have to set them up for every agent
- Thread handling: Snooze, Settle, Archive, Delete. It's the only tool where I don't get overwhelmed when working on multiple Threads/Projects at the same time (I am diagnosed with ADHD).
- Provider usage overview for all registered providers, showing pacing as well
- Great citation: when you make a selection in chat, a pill appears and you can create quotation elements that get added to that, can easily be edited or removed (the original reason why I even tried T3 Code in the first place, because it makes responding to long responses incredibly joyful instead of disorienting). I like it so much that I added the same mechanism to my own personal assistant/agent that I host on my VPS.
- Great diffing view where you can pick worktree diffs, or any diff from a specific turn in the current thread
- Has it's own subagent orchestration that's very pleasant to use
- Many more little things that remove a lot of friction and just feel right for me.
It's just baffling to me how it is better in every regard than every other harness that I've ever tried (Claude Desktop, Github Copilot App, OpenCode Desktop, all the CLIs, Integrated Agents in VSCode and Zed, etc.)
I will still start up a pi on the cli if I just have a quick question because T3 Code needs a bit to start up and connect to all providers/environments. This is my only point of friction: It's a bit heavy on resources (because it does a lot, not because it's slow. I just happen to use it on very old hardware and on one device that is very slowed down due to heavy-handed security scanning). But other than that, I can't imagine using anything else right now.
I don't know if I've ever been this exciting about a piece of software. I was very sceptical at first, but quickly fell in love with it.
I use Claude, glm , Kimi, codex and also outside of coding grok bot.
I’ve been building a harness and so use that and frequently benchmark against others.
I’ve also built tools so agents across accounts and across apps (Claude > codex > grokbot) can all communicate with eachother.
1) Claude Pro $20/mo (I'm retired now and just playing on personal projects)
2) I have API usage via Together for a number of toys I have built for my kid, I probably could refactor this but it predates getting a Claude Pro subscription.
3) Claude Code letting it do the work, pure vibes and dropping to vstudio as needed.
4) No
5) Muse, I have it doing things like basic research, waiting on ticket windows and just being a personal assistant. It also does a really good job at scraping churning forums and blogs and has made a few thousand dollars in the last month just on bank and brokerage bonuses. It does all the record keeping and I just sanity check the dates, hold periods and do the account opening / transfer work. I'm eager to see Anthropic's take on this space and hope it uses the same Chrome integration. Muse's biggest flaw is that it runs it's browser actions from a datacenter instead of a local host and a real browser.
Being retired in the age of AI is the best, all the time in the world to play with what really interests us plus an incredible assistant to squeeze every optimization path to productivity. Apps developed in minutes instead of months is like adding another century to our lives. And we are just starting...
https://github.com/seanturner026/f
There’s an excalidraw in the reader
My manager described this as incomprehensible lol but it’s modeled on conductor but uses Claude code terminal and mods and glues a bunch of open source stuff with a cli I had Claude write. Herdr for sessions, nvim for code review / searching, cli invoked by /slash commands. Beads cli for task tracking. With multi repo support since I end up doing a lot at work across repos at work for a “feature” which is how I delineate chats now. And I’ve got a nice picker for swapping sessions so screen real estate isn’t wasted. Many incremental improvements the sum of which are compelling ime
Phasing out my usage of cloud based LLMs/harnesses.
Focus is pi + self hosted GLM Flash.
Pi is lightweight and easily extensible and creates minimal extra noise. the output I get is very close to what I want to see.
my workflow is focused on clear turns between planning, testing and implementation. Any useful/recurring tasks, like getting/filtering logs, checking statuses etc get regularly folded into a project cli and agent skills that are local to the project.
I spend time interactively iterating on the plan/design and acceptance tests, then let it do its thing. Usually I have multiple interactive sessions at once that I'm steering this way. 1-2 assigned projects, 1-3 side projects (additional automation, experiments/POCs, workflow/process efficiency etc.). I generally don't do sub-agents. When I need throughput I automate first, frontloading the "deterministic" parts.
What are you using to host GLM-Flash?
Didn't set it up myself so my info will be meagre, but I think it's litellm and a big enough nvidia card. We use it for development but also for internal applications (where smaller models might escalate a task to it when needed).
Obviously not feasible for (just) personal use, the model is too large for consumer hardware.
I subscribe to OpenCode Go when I have a personal projects I need to work on. It's by far the best value since it's only $10 and I doesn't run out of its five hour limit in an hour they way Claude Code/Codex do. Models are more than good enough.
At work, I'm using Codex with the subscription they give me, and I don't really have anything to complain about.
I don't bother picking obscure tools and heavily configuring them since almost all of the intelligence is in the model, and models themselves improve so fast that any complicated setup will be obsolete in a few months. All of the useful behavior just gets trained into the models, and all useful functionality gets added into every harness very quickly.
I'm not using any personal agents since I don't see any use case.
1. 20$ Anthropic
2. deepseek api, local qwen/gemma on fw desktop
3. Claude & Pi
4. No 5. No
I'm basic "type what you want in prompt, read what comes out"
I use a tool I made that lets me orchestrate lots of AI agents (claude/codex/gemini) both on my laptop and on a k8s cluster. I set them up to run on various triggers and access them from my phone/laptop: https://github.com/jonwiggins/optio
1. Claude code, ultra plan, chatgpt go plan (India has a ~3$ plan) 2. Use pay as you yo openrouter with kimi2.6 for an api wrapper interface I have to work with my business APIs, connected to whatsapp. 3. I use opencode alot at my office that connects to a bunch of mixed models gpt-luna, claude. But prefer claude code, because (dunno if others feel it) opencode feels clunkier, and slow. I suspect that's because of the llm routing through my companies infra. 4. Have not made anything work, have tried GasTown, its horrible imo. Also, I think I am not able to trust the setup, and worry about token wastage (which cheaper models do more off when left unsupervised) 5. Open law - connecting with my personal calender, gmail, whatsapp, side gig infra - first responder for my business whatsapp for a slice of whitelisted interactions.
1) One 5x Claude sub, one pro lite codex sub
2) Not really, the value proposition on the subs are just so much greater.
3) I'm using Claude Code and Codex, crucially thru Muster[0] which I've been working on. It's a herdr inspired Libghostty-powered terminal which understands agents you're running, helps you manage them, supports remote devenvs, and it provides a CLI for agents to spawn, work, and message each other, no matter the harness and across machines.
4) I'm working on my own factory atm, which combines Muster and Kan[1] which is a offline, Kanban-as-plain-files-in-VCS project I'm working on and happens to work really well with agents. I recently added card dependencies to Kan and a kan ready command so I can have director agents orchestrate and hand out tasks based on what's ready and prioritized in Kan.
5) Not at this time. I played around with dots but I haven't been convinced it can do anything for me that my own solutions aren't better at.
[0] https://github.com/amterp/muster
[1] https://github.com/amterp/kan
>> [1] https://github.com/amterp/kan
Nice! I am guessing the intended use is on developers' own machines, not putting on a server and live co-editing in a team meeting. Pls. let me know if I have misunderstood. Thanks.
Yeah intended use is in every dev's own checkout of the given repo, on their own machine.
Models: I pay Z.ai a lite sub which costs me ~9USD/month and for the last months I've been using `glm-5.3-flash` amost exclusively for my day-to-day. I also make use of free models from OpenRouter for simple CI/CD automated stuff but very little so. When I'm out of sub tokens I fall back to DeepSeek credit. I keep my sub/credit usage public at https://a.l3x.in/usage
Harness: when building/troubleshooting I use https://pi.dev TUI most of the time, the rest of the time is via an issue/PR driven workflow also powered by Pi with this action I built on top of its SDK: https://github.com/shaftoe/awesome-pi-coding-agent/
My general take on AI at https://a.l3x.in/ai
Generally speaking, I think I can hardly be happier than what I am already, the quality of open weight models is incredibly good and the only upgrade I could think of is being able to run one on my workstation one day (I don't even try on my Macbook Air M1), which isn't something crazy to dream about, perhaps in 2/3 years running Deepseek v4 Flash or similar model locally won't require too much of an investment in hardware.
How often on do you hit the limits with the plan? Thx.
It depends very much on the project I'm working on, but recently I almost never run out of tokens before the weekly quota (which seems to be the tightest for me) resets.
I also use some discipline in not letting the context grow too much and usually keep every session very focused, context window fill rarely goes above 10% for me as I prefer to start fresh (perhaps from some markdown artefact saved from a previous session).
If you're used to throw every possible thought to an agent than most probably a light Z.ai subscription is not going to be enough.
I built railvio.com for myself initially, it’s taken me a year to build and it’s a full harness for building applications + fully integrated Jira and Confluence clone. That altogether has become a lovable clone which I’ve used now to deploy a tradingview clone, a risk rating rules engine clone, and an agent terminal that is a stripped down clone of Iterm2 but built with a middle layer so I can connect to long lived tmux sessions. 3x Claude ultra, codex pro, + local Qwen 27b fp8 on Radiance engine on 2xRS700 and ComfyUi on a separate machine with a rtx 4090. It’s been super fun building everything. I’ve averaged around ~40B tokens per month..
$DAY_JOB gives me GitHub Copilot with generous limits so I use GitHub Copilot CLI. We're a Microsoft shop so that dictates some choices too.
My workflow is a windows terminal app with multiple sessions running under wsl2 in each pane since LLMs seem to do better with bash than PowerShell. I don't trust these things much so I've got a careful list of read only tool calls and filesystem sandboxing set up so it can do some useful stuff like read my task board and publish draft PRs without a ton of overbearing permissions prompts.
I usually chat with that for a bit on a task until it's plan sounds good then I have it spin up a Docker Sandbox microvm under Windows so it can safely run yolo mode and make all it's changes. When it's ready I carefully review and iterate on the result until it's ready to publish the PR.
Bit kludgy but strikes a balance of safety and multitasking I'm comfortable with. IME most folks don't think much about isolation so curious to hear if anyone else has a similarly isolated but better workflow. I'd like to try OpenShell instead of docker sbx but it lacks microvm support for windows and I don't want another layer of nested virtualization since my windows machine itself is a cloud VM.
Work also gives me access to GHCP for personal use so I use it for side projects too. Idk what I'd use if I had to pay out of pocket, probably something a cheap Claude plan. For noncoding personal use I mostly just use free tier Gemini. If this stuff ever slows down I'll invest in some GPU nodes for my homelab and run local models.
Our token budget is also high so honestly I don't plan much. Mostly just throw work item and as much related context I'm aware of at whatever is the latest decent model (opus for a while and lately gpt sol) and let it do its thing. If I were more budget constrained I'd plan more up front and delegate bits to cheaper subagents I suppose.
I do basically the same. Have multiple Docker sandboxes where I let agents yolo.
Must say I don't really love the setup. The workflow is not as frictionless as just opening Claude Code used to be.
I am using mainly claude code on a Pro subscription.
I also have an OpenCode Go subscription, so I can keep an eye on the capabilities of the Chinese models.
Opus 5.5 medium for the vast majority of coding and code review tasks. I will sometimes save on tokens by dropping to sonnet if I have a well detailed plan built with opus.
I spend a lot of time on the design/architecture of features. These tasks are assisted by Gen AI but I still need to make a lot of corrections. It's my main bottleneck.
For orchestration and sandboxing I've used plain VMs (I have a make VM script which installs Claude in a fresh VM).
I've used Scion at work, which gave me a lot of ideas and got me excited about building a better orchestration, messaging, sandboxing and web UI thingy which I am currently working on (GitHub link in my profile).
My biggest pain point I wanted to address with my tooling is cleanly separating projects and accesses. I do a lot of web dev and sys admin stuff. I'm building Egzo so I can safely grant and monitor access to remote systems and cloud services and keep myself organized.
For my side quests at any given time I am subscribed to one or the other of Claude $200 or Codex $200 or both. I was briefly on 200+500. Getting to be a real expensive hobby.
At my day job I spend about $1000/mo at API prices. So like 5-10% of my hobby usage.
Hobby setup is I have a single "personal assistant" session which spawns sessions for each project or major job on a project and messages with them. I talk to the personal assistant over Signal exclusively.
Work setup is just a wrapper around Claude Code.
Inference: combination of Opus 5.5 (best model in the market) and local models (DeepSeek 4.1f 1017 and GLM 5.3f) running on 3x DGX Spark.
Harness: Primarily Hermes Agent across desktop, terminal, and phone (via Telegram). I have also tried OMP but find it a bit clunky to use. Hermes tui + tmux are my favourite tools and Hermes works well with local models.
Orchestration: I haven’t seriously worked with it yet, but starting to use herdr with Hermes. That is as much “software factory” I get.
Claude Pro. Mainly used for developing my apps Pagecord and Feedgrab, but also for personal stuff like diy problem solving. travel planning, data chores, etc.
I feel I get a wild amount of value out of £90/mo (paid for via my apps revenue)
1) 3x Claude Ultra subscriptions and $20 tiers of Codex/Cursor/Gemini, the latter three mostly for multi-provider testing and Google's bundled cloud storage. I basically only use Opus xhigh, but never touch Ultracode.
2) No. The subscriptions are too much value compared to API. If i were to start using any model on API, the first one I'd consider would probably be Inception Mercury 3, for the 1000t/s speed.
3) I use my own custom kanban agent management tool named Kotkit, which I have been developing since March. It has a built-in multi window desktop which I can use from my phone as a PWA.
4) I am currently building orchestration/swarm capabilities into my tool. Currently adding multi-agent deep research and code review workflows orchestrated by my tool, so that they are cross-provider and my agents can trigger these on demand, as well as experimenting having one agent complete an epic ticket on the kanban board by orchestrating task agents (this is all wired up so AskUserQuestion and ExitPlanMode, etc. on the task agents are answered by the main agent, and if the main agent is unsure, it delegates to me). Currently doing it on a >20-task code migration epic with a lot of success.
5) No. AI is not yet good enough at long-term planning for these to be useful for real work. You have to build step-by-step up to it. These agent platforms are good for simple personal tasks, but to use them for actual work would be skipping a lot of steps.
1) i use my brother's claude pro subscription. if i needed to, i would use openrouter + API if i didn't have his subscription, but i don't think i really need it. i've never purchased an ai subscription myself, nor paid my brother for his. i would never use frontier us models for anything serious on openrouter + API
2) yes, when claude starts getting too preachy and the task at hand is important to me i use openrouter + api + chinese models. my computer can't even run qwen 3.8 27b.
3) i use claude code because i think using other ones isn't allowed. i don't really like it, and generally prefer opencode (i've had to use it to do some things for my brother). i'll try using other ones though
4) no, and i didn't actually know they existed. opencode and claude code's agentic things are enough for me, though i don't do actually complex projects
5) no, i won't use them unless i get good local inference going (i.e. never).
1. Claude Ultra (200) and ChatGPT at 20
2. Experimenting with qwen3-coder-next 80B and a few other local models for coding tasks on my M4 MBP (128GB RAM) - no other models or pay-as-you-go apis
3. I have a custom built setup primarily driven by claude code via telegram with a checked in "brain", a separate "watch" structure of crons that i can setup and tweak, and then a task farm run by that orchestrator with containerized cc (and a mixture of sonnet and local llms based on task) and application devcontainer stacks (if it's coding related)
4. Custom
5. Custom - mainly rebuilding the ideas from there into my own workflows, but not using any of them directly
Trying to split more things out of CC and use it as a pure harness on route to trying other harnesses out like pi, etc -- it's a work in progress as and when I have the bandwidth.
$200/ea plans on Claude and Codex, plus OpenCode Zen for experimenting with other models, plus Qwen3.8-28b locally. (I test-drove DS4 Flash locally in a 128GiB MBP for a while, but the tokens per second were too slow to be useful for more than being a toy.) At my prior anchor client I had an additional max20 plan; at my current anchor client they have a teams plan.
I use both Openrouter and OpenCode to get Kimi, GLM, big Qwen, etc, but not at great scale. I have a security product that consumes API tokens that I built against Opus 4 when that was the strongest available; when I get around to it I’ll likely shift it to prefer GLM or Kimi, slightly for cost but mostly for non-refusal.
I spend a lot of time in Claude Code because it’s the only Anthropic subscription-friendly harness. Additionally use OpenCode and codex. Not much time spent in pi these days. I used Aider before Claude Code’s beta. Claude Code’s plugin framework is pretty decent, so I tend to stand up a repo of custom plugins against every client repo I work with. It’s a nice efficiency layer for me up front and often grows into a bonus client deliverable after a couple months of refinement.
OpenClaw runs persistently, mostly for little calendar tasks and trying to mitigate my ADHD: energy level check-ins, redirects to use attention for things that matter. It orchestrates Hermes under the hood but I don’t interact with the latter directly.
I still mostly just use tmux directly and tell running sessions about each other, sending messages to each other via text files or send-keys/capture-pane. Herdr is installed but mostly unused. The GUI clients aren’t my thing; I’m a born console junkie.
What I really want to change next is pulling this all back into emacs.
1. none 2. nvidia nim, opencode zen, google free models. local QFN iq3_xss on m1 max 64gb, paperniuk splash 3. pi agent, pi pocket 4. no 5. no
For local inference, DS4.1F 1017 quantized to 2.25 bpw from oh-my-pi; this is slow (~500 tps prefill, 30 tps decode) because it's a Radeon R9700 on PCIe4. When I don't need local inference, Opus 5.5 reviewed by GPT-6.1 Sol in VS Code, both $20/month. Mostly I use them to build small analysis tools because I am not really an individual contributor any more. I used to hit both my session and week quotas frequently, but that is much less common since Opus 5.5.
1) Claude Ultra (aka Max 20x)
2) No
3) Claude Code
4) Custom WIP https://muster.town/
5) Custom personal agent setup https://github.com/mitchellgordon95/miniclaw
Usage is mainly wiki building around personal info (finances, topics of interest), calorie counting, and doing various things from my dumb phone (rideshare booking, checking MTA Bus schedules, etc.)
1. I use both Melious[1] and Entrim[2]. I have a Melious Pro subscription, €150/month and use Entrim pay-as-you-go API, because it's cheap.
I used to have two Mistral subscriptions, because I needed the extra capacity, but quickly realised I'd need three or more, so switched to the two providers above.
2. Providers covered in the first point. Regarding models I use a combination of GLM5.3, GLM5.3 Flash and Deepseek 4.1 Flash 1017. Which model I use depends on the task. GLM5.3 Flash for general chat and co-ordination tasks. Deepseek 1017 for most coding related work and GML5.3 for adversarial batteries and documentation.
3. Maki[3], although quite customised with various plugins. No python permitted, only Janet for scripting. Extremely limited bash calls, instead preferring to use the built in harness tooling, or plugin tooling I've written. AGENTS and skills used extensively. Mechanical gates also used to prevent certain actions, like remote git pushes without explicit prior approval.
I switched to Maki for the token saving strategies and stayed for the ergonomics, which I think I like, but maybe just feel comfortable with now.
4. None listed. The closest I come is a multipane Zellij setup with different agents running in different panes. Orchestration is still human in the loop though.
5. None
[1] https://melious.ai
[2] https://entrim.ai
[3] https://maki.sh/
For personal use: Claude Pro, Opus 5.5, occasionally API / usage credits. Outside of that, some smaller local models (lately poolside)
I still use emacs, I run agents in buffer or in different terminal tabs. I like pi but only claude is approved at work, so I mostly stick to that for consistency.
I'm use 1-3 agents at a time, not more than 2 on the same branch. I have some pipelined workflows but I don't run swarms.. I'm on budget!
Claude Code Ultra aka 20x Max, about $250/mo incl VAT that I have to pay
My challenge is revisiting AI-gen code months later and making sure the AI has the context from our previous work.
How do people handle this?
Do you keep everything in the same conversation thread?
First off, I cannot pretend I have this solved. But I can share one thing that has worked for me, along with codex maintaining it's own local memory (turn it on).
Have your agent maintain a task journal. It is like a more detailed commit history, that includes tribal knowledge gained during the task and additional context, eg.
Reset your context frequently instead of only relying on compaction. Use the following prompt to summarize the current context for the new agent: Give it a try and feel free to share any improvements.You don't.
The context should be self-evident from the code alone, and if it is not, it should be in your head and you should write it down again so that you do not forget it yourself.
Being forced to re-explain stuff to the LLM is good, actually, as that way, one actually thinks about the architecture, choices and limitations.
Try to keep context shoved in a conversation thread under a token limit to save tokens.
Ask AI to record in a projects directory memory: LOG.md, HANDOFF.md, DECISIONS.md, a plan folder. Reference code with absolute paths. Keep repeatable processes in skills. Version control it all, and use a dreaming process across multiple sessions to stop repeat mistakes.
That's what's worked for me, but there's many variations of this.
> Try to keep context shoved in a conversation thread under a token limit to save tokens.
For big tasks/features, I'm always creating a new git worktree and I'm using a "status repo" skill that creates a sibling status git repository where everything is recorded.
It works quite well whenever I have to revisit something. It's quite token hungry but I have the full trace of what happened in the past.
https://gist.github.com/vkalintiris/3ad526e71a75bcba013aaadc....
Another approach I want to explore is using https://github.com/netdata/netdata (disclaimer, I work at Netdata) to send all OTel logs/traces of LLMs to it. The idea is to replace the status repo skill with something that will allow LLMs to retrieve the history of a given task/feature. It will probably require some "checkpointing" to make this more efficient.
^ This is basically what I was getting at. Thanks for putting it so clearly mate.
> Try to keep context shoved in a conversation thread under a token limit to save tokens.
That is another solid tip. If using subagents, use the main thread for only the conversation + launching subagents. Tell it explicitly not to implement anything, and only support integration efforts.
Doing so keeps the worker subagent's context clean - so it can stay laser focused on the task.
Have your agents keep contexts of tasks up to date as they work on stuff.
I've been iterating on this approach for about a year—it makes switching between agents and harnesses trivial.
https://photostructure.com/coding/claude-code-tpp/
I usually have a “close out” routine that’s something like telling it to record hints to the project’s agents.md. It’s usually pretty good at figuring out what to save, and I usually set up the prompt to point the thing’s attention at the part of the process that made me most frustrated.
1. Claude $20/month and Cursor $20/month (I've been meaning to close it) I mostly use Opus 5.5 on Medium. I've been experimenting with High effort, mainly at work.
2. I have an OpenRouter $50 pay as you go API key that I use for Mercury 3 at this point, primarily for tokens per second. I also use GLM 5.3 when I inevitably get [cyber] blocked (or for reverse engineering in general, apparently) by the morality police at Anthropic.
3. I run things pretty stock. OpenCode with OpenRouter but otherwise I just use Claude code.
4. I've never used these, but I'm curious about them at this point.
5. Not yet, but also interested. My plan is to use Hermes for some personal agent work on a local model, but I have not set it up to test this yet.
1 - Opencode Go $40, Claude $20, $0-100 on Deepseek API
2 - No
3 - I use a lot of Opencode + Deepseek, strategize some of that work with Claude Opus, and some gtm, research, and project stuff in Hermes with its coding agent. When I guide by hand I use Orca Dev, which is lovely.
4 - Claude drops GitHub issues that Opencode picks up on a schedule. It’s simple and resilient enough for me. Hermes has a lot of scheduled tasks to move things forward.. it has gtm and research and dev loops (the latter builds autonomously for future projects vis a vis my attention, I’ve used quite a bit of it).
5 - Hermes Agent, which has become a core tech for me. I’ve tried to use Openclaw but it has not been something I can keep running. It spirals and crashes after a week in ways I’ve never experienced from Hermes, doing the same things. I have little interest in those generalist agents, maybe after a year.
I'm currently using https://www.conductor.build/ for working on multiple projects at the same time with git worktrees and my Claude subscription. The annoying part is that if I have a frontend repository and a backend repository I have to tell each session "talk to this workspace to adapt the backend" if I'm working on the frontend. I wish they had better support for multiple repos in one session.
Separately I'm evaluating https://ampcode.com/ which is very polished, but I feel like the diff-reading part is not very prominent there which I find a bit odd. After watching many of their videos I feel like that's because they are opinionated on this not being necessary any more as the models are good enough. I kind of agree, but still like to give it a last glance before committing.
I ended up just making a monorepo because it makes using AI agents so much easier.
Claude Code (and Codex too I believe?) can message other sessions these days. I just run one session per repo and they talk to each other directly when the frontend needs something from the backend and vice versa. This has removed most of that back and forth for me and eliminated the need for orchestration tools.
I use Claude, on a Max plan.
Have a server at home, so mosh into it and run Claude Code at the lowest level. Use Blink off the iPad.
For normal work though, I have a bunch of services set up to run Claude —-remote-control in every one of my projects, so I just use the Claude app on my phone and iPad.
Also maintain and run https://github.com/common-pattern/claudeflow as a service the bigger projects - it’s an orchestrator that reads GitHub issues and starts a Claude one-shot, automatically handling the routine stuff like CI, commits, PR landing flow etc.
1) no personally paid - employer gives me claude 2) no 3) corgi + claude at work 4) no 5) no
For building, I started with Kiro and later switched to Antigravity (Opus 5.5 for planning, Gemini 3.8 for implementation - but now using Argon for everything and quite happy). Claude for design.
Had a Hermes set up with a dual NVIDIA DGX Spark and Deepseek v4 Flash, but then built my own harness (getcinnamon.ai) and pretty much switched to that 100% including family members.
I use a nix one-liner to run llama.cpp on M5.
``` nix run nixpkgs#llama-cpp -- server --models-dir models -m models/gemma-4-12B-it-Q8_0.gguf --port 8080 ```
open http://localhost:8080.
I also run pi in a docker container to do some vibe agentic coding. But I prefer `cat flake.nix | pbcopy` and paste into the browser. Then I can export the jsonl files and use tools to convert them to md and push to GitHub Pages.
I can do most of AI sessions off-line and only need hotspot to sync with GitHub.
My main use case is to learn and hack nix on Mac. It is fun.
I use Muse Code $5 month
It has a 5 hour usage limit, and if using the wonder model, you basically only get 1 or 2 prompts. However, I find it will basically do unlimited 'turns', so as long as I give a good prompt, I can get my work done.
No ide/harness. I only do light coding and don't have time to actually sit at my computer.
1) None.
2) Yes, trying different models and harnesses through many providers (free/trial). mimo-v2.6-flash, minimax-m3.1-flash, glm-5.3-flash, deepseek-v4.1-flash, gpt-oss:120b, nemotron-3.5.-super-vl.120b-a12b, and others.
3) Mostly eca.dev on Emacs, and Claude Code at work. Now trying the OpenCode one too.
4) No.
5) No.
1, work - claude team pro (not paying for), home, codex pro (for side hustle https://rolloutrhino.com/)
2, no, except at work where I have a whole agentic helpdesk system that self improves, creates documentation, and is quite awesome really. Pretty cheap to run also.
3, my own that I keep iterating on, but I do use the clis, though with codex, the app is just awesome, don't use the tui (its good, but the app is better). Claudes app sucks in comparison, though is often better than the TUI.
4, my own tools, mostly because I started on my own ones ages ago.
5, dots, It's kind of like a project manager for the side hustle, its really coordinating a lot of stuff between all the things I have in flight
Curious what you find better about the codex app vs the claude app? I use both as well, but I find them fairly interchangeable. I like project and session creation a bit more on codex.
My own agent: https://github.com/fluxmoot/fluximp
I primarily use DS v4.1 Flash @ deepinfra.com with less use of GLM 5.3 Flash and GLM 5.3 @ friendli.ai
I use a locked down (i.e. many tools disabled; shell especially) setup of oh-my-pi in benchmarks for comparison.
1) $100 Anthropic via work 2) codex personally 3) Claude Code + Herdr with a root agent steering everything. Waiting for Rex 4 and 5) absolutely not lol
These comments read like people just bragging about having multiple AI subscriptions but not actually giving any info about what they're using them for
I can't for the life of me figure out why people need to spend so much. Maybe I'm just using them wrong.
I think a lot of people/companies are using 10x the tokens for 10% improvement. If you're willing to spend a modicum of attention/effort looking at what the bots are doing, one max Claude plan with Opus and occasionally Fable is very effective.
I like using Cline the VS Code extension. It can use local models and local inference with Ollama and tons of external inference providers like AWS Bedrock, Vertex, etc. It asks for permission to read webpages or run commands. You can give it whitelisted permission to just read your files which is what I do. If you want, you can put it into ACT mode and give it permission to write to files but I avoid that. If you use something like Bedrock, it shows you how much the conversation costs. It works great with Claude for coding but if I'm doing something super small I might use a local Qwen or Gemma model.
I use Claude and codex and use my own orchestrator to run them. I also used to have open code for a while but the recent codex and Claude models are token efficient so don’t need additional models.
I use Circus Chief to orchestrate workflows that use various combinations of Claude, Codex, Muse, z.ai, and Grok subscriptions. When one subscription runs out, Circus Chief falls back to the next. You can configure tiers to control how that works.
Full disclosure: I’m the creator of the project.
https://github.com/ferrislucas/Circus-Chief
I'm a solo developer. I'd hit walls with openrouter (sonnet/opus/sol) + cline
Now I'm running qwen 3.8 FN + omp + https://github.com/mattpocock/skills
This combination has pretty much "solved" node/react coding for me, running locally on dual RTX 6000, 200-800 tps depending on concurrency.
An incredible time to be alive.
Deepseek Flash or GLM flash, paid upfront using Openrouter, via Cline on Visual Studio Code.
I've been increasingly interested in language-oriented programming and denotational design. So i get claude to create a domain specific language in Racket modelling the software i'm trying to build, we create nouns and verbs whose meaning fits the behaviour i'm going for, then when we find bugs or notice we need more exceptional (less-elegant) one-off behaviour to tackle edge-cases, we iterate on the language in such a manner that it these are tackled emergently with preferably less concepts, more cohersion and composition. It's somewhat similar to how memory safety is handled at compile time in Rust but for business logic.
This sounds really neat. I'd love to see an example.
I am currently stacking 5 opencode go accounts, spending $50 a month for $500 in deepseek v4.1 flash credits, using OMP as my harness
Copilot in VS2026 + GPT5.5 configured as a custom model (API Tokens). Same as it was 6 months ago.
I really want to use 6.1 Terra, but the way copilot is set up in visual studio proper, they don't handle reasoning models correctly (or maybe my install is screwed up). Eventually, I will dust off that old winforms project and get my custom harness working with the latest responses API contract.
I prefer (pre) paying for tokens. You can offer me all the proof you want but I'll still never believe that those subsidized tokens are always the same as the ones I get through the API.
Claude Pro for personal. (Kiro at work.)
Moving towards long-running, asynchronous delegation where I use a locally running instance of Fizzy kanban board to manage and track all work.
When I think of something I might want an agent to do, I create a card. When I want to delegate I assign to agent and move to To Do. I usually try batch things, then use a slash command to invoke a fizzy skill. I do this in a terminal and then close it and do other things.
Couple hours later I'll check the board and see what's in review column for me. Read comments, look over PRs, etc. Agent has broad access to specific project folders and repos.
In general feeling more like working with a colleague than chatting with a bot.
1. Z.ai coding pro plan that i got last christmas for $14 a month. Will not renew in jan.
2. Locally hosted qwen 3.8 flash next on a wake on lan machine rtx5090
3. Currently orca as front end orchestrator to a mini pc with pi for coding and hermes for general purpose
I'm currently on Raven as a sort of "meta harness". The main chat is with the orchestrator, which delegates all work to other harnesses. I just use different seat-based plans. Mostly the latest Sol with my corp chatgpt account.
I get a lot of value from having an LLM to talk to LLMs for me. It can automatically set up adversarial reviews from other ecosystems, or just prompt the implementation LLM to adhere to certain standards. My goal is to have fewer touch points when implementing. I want to speech to text a paragraph or two giving "taste" to the process; everything else is just toil.
What is your experience with Raven or such meta-harnesses, does it consistently outperform vanilla?
I get the same "one-shots" from the vanilla codex as codex under raven. It took some work to get the orchestrator to send good refinement prompts, and it still doesn't have good taste. I need to bring that with my own review of the code. If you notice you send a lot of repetitive refinement prompts it's worth it.
For software development, I am using Kilo Code (https://kilo.ai) extension on VSCodium, with Kilo credits that gives access to hundreds of models at API pricing.
For day-to-day tasks, I am using Goose (https://goose-docs.ai) with OpenRouter.
1) Claude Pro (the $20/mo tier). No other subscriptions.
2) I use models on open router occasionally, primarily GLM and Deepseek models. I’ve experimented with Qwen locally but the models that run on my hardware are too dumb for extended use cases and too slow for quick questions.
3) Personal use: Primarily Zed. I use employer paid Claude Code in the day job but it eats tokens at a rate I’m not willing to pay for for personal use. I’ve tried OpenCode but their security track record is uninspiring and I’m way more defensive about my personal hardware than my employer is about their hardware fleet.
4) No
5) No
1) Work covers my Claude and Codex subscriptions. Each is the ~$200/month tier.
2) I have a "live Zoom call audio transcription" setup that uses Silero VAD, Qwen3-ASR-1.7B, Qwen3-4B-Instruct, and pyannote speaker-diarization-community-1.
3) Claude Code in the Desktop UI using remote control environments.
4) Custom orchestrator for having Codex review Claude or vice-versa: https://github.com/navels/neal
5) I have built over time a personal assistant using Claude remote control running on a mac mini, this is likely similar to Dots?
1) I bought a z.ai 1 year subscription for their mid tier (second iteration of their pricing) for like 500€ total, so 40€/month for mostly GLM-5.3. I suppose flash is there too, but.. eh.
2) I have a 4090 which runs Qwen3.8-27b q4 fine with 180k ctx at q8. The model is capable enough to do things, but dumb enough to not be able to build elaborate nonsense like the bigger models love to do
3) I'm using https://github.com/juggler-ai/juggler which is alright. It could be a bit less vibecoded (and the readme and docs and all would benefit from less LLM writing), but, generally, the opinions shaping it seem decent, and I like just how much control over and insight into what happens it gives you.
I initially tried OpenCode, and, holy hell, what a mess. Do not ever use OpenCode. I also tried goose but it was too agentic for my taste.
4) No. I am doubtful that that kind of workflow can produce software I would deem worthy of production. I also personally just dread the whole overhead that is tickets, jira, etc, so recreating the same but now you're the manager seems like a personal hell to me.
5) No. I wouldn't know what for
1. max 200 on claude (3x) and codex (1x) 2. yes, mistral / glm / deepseek via openrouter and some others 3. custom 4. custom 5. no
1. Claude Ultra
2. No
3. Claude Code
4. My own engineering harness: https://devarch.ai/
5. No
I started using ChatGPT in the beginning then switched to Claude then Claude Code then developed DevArch over that entire time.
I only use genuine intelligence
Same.
For 4 years now I’ve tried the latest and greatest once annually before my summer course starts— just to know what to expect and how to instruct. The artificial stuff is definitely getting more refined, but using it doesn’t yet make sense for me personally.
I am committed to learning about it; it would be nuts not to.
Not committed to using it.
Until recently my mental health post-burnout has been so absolutely poor that I have been learning all of this just to try to concentrate on anything so I don’t feel further out of the loop. Most days I still feel like I should quit IT.
In the limited ways I am actually using it in a productive rather than self-educational sense, I tend to be using it to combat burnout: helping get past blank page problems, rough ideation and keeping track of ideas, doing slightly more cohesive research than just ending up with dozens of open tabs and some unclear notes, generating helper processes (like content migrations) that I might otherwise not be able to make myself write, and exploring using it to do the unthrilling work like automated test generation.
I don’t plan on becoming someone who sits back and specifies loops, and that doesn’t fit with the work I am doing. I have far too many disjointed things I have to work on.
But it is definitely a more varied rubber duck, and that has some potential. I feel a little less stuck, sometimes.
At work:
Opus 5.5, fast mode, xhigh/max reasoning, ultracode as needed. Usually 5-8 agents going at any given moment. In-house frameworks for Claw-like workflows and agent orchestration.
At home:
Cry
1) claude, claude again, codex, antigravity, cursor 2) no 3) using a variety of different harnesses managed by codealong 4) built https://github.com/codealong/codealong 5) openclaw not used very much anymore in favor of native harnesses
so much inspiration in here!
hard to contain myself not to change my setup too often (instead of doing actual work)
Claude Code Ultra in Emacs. Split Emacs windows: Bottom is Claude Code, top left is code, top right is db-backed editable spec (propositions and premises) with Clojure like eval that sends spec to Claude Code.
My workflow has simplified a lot over the past year and I personally haven't seen a need for complex orchestration tools or the agent platforms like OpenClaw / Hermes.
Since agents can message each other nowadays, most of the time I just talk to one Opus 5.5 session (because of its 1M context window) and let it talk to all the other sessions, acting as an orchestrator that I use from the cli and the mobile app. I also have some of my agents reachable over IRC through a small CLI I'm building.
I subscribe to both Claude and Codex. Mostly at the top tiers, but I downgrade every once in a while based on how much I'm using them.
I don't use Codex directly as much, even though their harness has gotten a lot better and snappier over this year. Most of the time Claude is the one using Codex on my behalf. I was using this delegation workflow a lot until Anthropic fixed the word salad issues on Opus 5.5. Nowadays I mostly do it to load balance my token usage.
I also use locally hosted models (both via CUDA and MLX), but much less for coding unless it's for asynchronous tasks that I can't use subscriptions for. My main use case for local inference is privacy and tinkering.
I've also been building my own harness focused on security and privacy for the past year or so, but it's still a work in progress.
1) Neither. 2) Yes, I use pay as you go from OpenAI and Anthropic. 3) OpenCode 4) No 5) No
I use OpenCode to update my Home Assistant automations. Home Assistant also does two lightweight checks a day using gpt-5-luna (Look at this photo is the pool cover on or off?). I also use OpenCode for a few other small basic tasks.
This is me personally. At work I have a ChatGPT sub an a, seemingly limitless (I don't use it much tho) OpenAI API key.
two max Claude suscriptions, active dev in vscode to share screenshots, GitHub tickets for memory, basic skills, superpowers,subagentic coding with opus and the main in fable, daily Cron tasks for answering email/linkedin/etc. Desktop Claude for video/music editing . No API usage, expensive for my needs.
Locally (MBP 128Gb) I have MLX-Serve constantly running with Clef (for model routing and decisions), Qwen-35B-A3B and Ornith-1.5-35B.
And I switch to Qwen3.8-27B (the Perplexity Pro version that works well with MLX-Serve) or GLM-5.3-Flash (sushi) when I need to do serious work offline.
As a general harness for coding, etc. I use Swival.dev . Configured to use Clef for model routing, with Qwen-35B-A3B for documentation tasks, Ornith to ask questions about code, and various GPT models from a ChatGPT subscription for more difficult tasks.
For music, pictures, etc. MLX-Serve is all I use, with the recommended models. It's really nice.
And as a general agent for research, etc. I absolutely love Perplexity. This is by far the AI produce I use the most, ever since it's been available.
It's a little bit sad that they are forcing people to use Perplexity Computer, because it's so expensive. But it works incredibly well.
Perplexity can now run locally, but even though they include the best Qwen3.8 quant in my experience, the Perplexity tool itself has major issues with context management and tool calling. So, when that doesn't work, I fall back to MLX-Serve and Swival with the Chrome MCP server.
For agentic web browsing, I can't live without Comet anymore. So many websites are painful to use, and finding the right buttons to click even to perform very common tasks is a challenge. So, I frequently just ask Comet to do it for me. I also use it to compare products, compare places on AirBNB, write personalized messages on Instagram, etc.
As for an OpenClaw-like tool, I'm using NBClaw. I was initially very excited about it. Look, I can just send messages on Signal and let the agent do work for me! And I occasionally found it useful to remotely fix issues on the host it was running on, or long tasks such as making sure that all my projects worked with a new version of the Zig compiler. It became a replacement for Perplexity because, hey, it's cool, everything's running locally, using my own web browser, etc.
But eventually, I used it less and less. Nowadays, I still enjoy the fact that it's sending me a summary of cool things it found in Hacker News and research papers on topics I'm interesting in. But that's pretty much it. I don't use it much for agentic tasks anymore. It still happens (I recently used it to fix a bug in an opensource project while I was on the bus), but mostly just only when I ran out of Perplexity Computer credits.
I was using VSCode with Claude Code extension, multiple tabs at once, multiple instances per worktree.... then I listened to the Lex Fridman episode with DHH [1] and they mentioned Herdr [2] and I tried it out - NEVER LOOKED BACK! It has been an extremely long time since I completely abandoned and changed my setup - surprised myself.
Herdr is awesome - it pings me when an agent is done or needs my attention and shows me all my terminal panes that are running an agent that I can access with one click.
I am now 90% Herdr terminal, open IDE maybe 1-2 times a day just for more convenient way to look at something. I searched ways to stay in the terminal - found lazygit and love it!
I use long-running worktrees (e.g. named worktree-1, worktree-2, worktree-3, ...) and have parallel work going on in them so I can work on whatever I want.
In each Herdr workspace, I have three tabs: 1st is for ~3 panes, one for Claude and two or more to run the app and basic terminal. Tab 2 for lazy git, and Tab 3 for an extra terminal when I need one. For example, right now I have ~20 Herdr workspaces open now between a few projects, multiple worktrees per project and a few across remote machines ready in case I need them. When in the zone, I might have 1-8+ agents working on completely different tasks.
I have a couple remote machines over SSH that I use when I need a task to run and I need to close my laptop.
I have a keyboard shortcut to init my workspace to that setup. lazygit is amazing.
I ask claude to do everything for me, for example: "Let's do it, open a jira ticket, assign to me and put into sprint. Fix it and open a pr"
I create artifacts all the time to share findings with team or as a checklist of manual testing I need to do.
I find that gh copilit does pretty good review on PRs, and I have a claude skill to go address all the copilot feedback (review, validate, reply, resolve, push changes if any were made).
I was on the 200/mo Claude plan but had to switch to the Team plan, so I pay 120/mo for two users and I swap to the 20 plan when I hit my 5hr limit on my 100/mo plan.
[1]: https://lexfridman.com/podcast -> DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux [2]: https://herdr.dev/
1) None.
2) No.
3) None.
4) No.
5) No.
This is the way
watch_out_were_dealing_with_a_badass_over_here.jpg
My halftop macbook m1 max 64gb is doing all my ai locally, i have not paid for any ai subscription for over a year.
things i run locally using apple mlx models: general llm, coding llm, text-to-image, text-to-video, image-to-video, text-to-music,
everything is holding up well. I may eventually swap the m1 max for a higher memory bandwidth device once they are available for a reasonable price in 6+ years
Can you explain how you use a local llm on your Macbook Pro to someone who has no experience with LLM?
1) I pay for claude ultra coz the plan rates are significantly better than the api rates, and i feel like i can tokenmax harder than if i am constantly watching my credit balances.
2) yea openrouter free models, usually nemotron coz they have decent availability and I don't get constantly rate limited like with the gemma models on there.
3) honestly still using the default IDE, used to use claude code cli but nowadays i just launch claude remote control and just talk to it through my phone.
4) something custom, i wired it all up in slack so i could easily talk to agents through the same place i do my work
5) Used to use openclaw on a raspberry pi to experiment with like autonomous development but i needed something with a bit more transparency so i moved everything into slack.
1) What AI subscriptions are you personally paying for?
My employer pays for copilot, business plan. It also pays for Gemini.
2) Are you using models outside (or instead) of those subscriptions?
Gemini, company-wide. I mostly use it for general research, making clip art, and occasional programming questions / technical questions. I don't use it everyday, but when I use it, I really like it.
I no longer use ChatGPT or Claude's general-purpose chat bots.
3) What coding harnesses and interfaces are you using? Are you sticking with the default CLIs or IDE integrations, or using alternatives like OpenCode, Pi, Aider, Muse Code, or custom-built harnesses?
IDE integration with Visual Studio, and Copilot code reviews on Github. I mostly use agent mode; sometimes I use ask mode, and sometimes I use plan mode. (One very important thing is to instruct copilot, in .github\copilot-instructions.md, to only use its edit tool and not use Powershell to edit files. Otherwise you need to babysit it and approve cryptic find and replace commands.)
For Copilot in Visual Studio, I almost always use Claude Opus 5.5. Some of the GPT models "appear" to work well, but they leave the code harder to maintain.
We now do a round of Copilot reviews with our pull requests in draft mode before asking others to do the review; and we're not spending as much time "reading" code when looking at pull requests.
We don't (yet) have agents running unattended with high-level instructions and then making pull requests.
4) Are you using any agent orchestration or swarm tools?
No
5) Are you using broader agent platforms like OpenClaw / Hermes Agent / Grok bot / Dots / Muse?
No
------
One important thing: I'm after my boss to dedicate time (1-2 weeks) for me to set up and learn more of the hands-off tools. I love working with co-pilot, but it's best for work where I need to be hands-on. I really need to be able to offload simple work, or work where lower-quality results are okay, to a bot that I don't need to supervise.
I feel like this is a great opportunity to mention https://github.com/regnull/openbot, my attempt to create an open-source tool similar to Grok Bot. You can hook it up to OpenRouter and solve 80% of problems with 1/10th of the cost. I use it all the time. Contributors are always welcome.
1) Right now I have Opencode Go Plus, Mistral Pro, Kimi. It changes. Kimi is not staying and I'm probably getting ChatGPT for a bit, last month I had Claude Pro. I stay around €100 a month total.
I generally use open weight models (Kimi K3.1, Mistral Large 4, GLM-5.3, DeepSeek v4.1 Flash, Mimo v2.6 Pro is the current quiver) but I use closed frontiers enough to stay up to date.
2) I use OpenRouter if I have a specific need, audio transcription (MAI-Transcribe 2) being something I use often enough. Also using it just to test random models.
Not much local anymore except classifiers / decision models that I've played with.
3) Pi all the way.
4) I use Paseo to orchestrate agents, I have a couple plugins I've written for it that manages pipelines and such. I generally start a workflow with a simple skill and if it proves useful I like to make it deterministic if possible. Models be dumb sometimes.
6) No but I have some Paseo plugins for personal use, though it's very limited what I find useful. I also don't find that I want any claw type stuff at all.
1. Personally? Claude monthly and I also keep money in the API service and slowly chip away at it with various apps and toys (I created a Fantasy Premier League app that does research and sentiment analysis as part of its runs and that costs money... but not much) . Professionally I use an unlimited Claude via AWS Bedrock.
2. I use both qwen 3.6 35b-a3b and qwen 3.8 27b on my Macbook Pro (M5 Pro, 48GB, 18/20). I use those with ollama and hermes (and Obsidian, to an extent)
3. Hermes Agent, Claude CLI and Claude in VS Code
4. No
5. Hermes Agent
I use a carefully curated selection of agents for my professional work. Agents to handle adversarial code review, planning, questioning, dev work, etc. I also have custom instructions for various repos that need them (where I can define things like repo norms that are not immediately obvious, data models where needed, etc.
Mostly Pi, using a variety of models, but mostly GPT 5.6 line of models.
I try to use open weight and cheap models as much as possible.
I'm using claude with a pro subscription. Claude Code extension in VSCode devcontainers and Claude CLI in a restricted Windows account in powershell terminal for android development. No orchestration other than a directive in CLAUDE.md to allocate coding tasks to sub-agents.
I've moved all my ssh keys into a key-agent which prompts for key use as I got alarmed by claude attempting to login to remote systems to verify things as it worked through problems.
try https://www.droneship.run
Take a look at gnosh.ai; it is a frontier harness worth understanding. You will hear about it in the coming weeks.
I have two $100/mo Anthropic accounts (watch out the single $200 doesn't raise weekly limits) and one $100/mo Gemini account. I have Opus generate review prompts for Argon and paste them into agy.
I tried setting up a script so Claude could automatically call agy, but it kept hitting timeouts, and letting Claude control that tends to burn my usage too quickly anyway (and if I'm not copy-pasting into different terminals, where's my value add in all this?).
Argon hallucinates, so I rarely use it for impl, but I've actually had better results with it than Opus for SwiftUI and AppKit work.
I try not to run more than three sessions at once, or I get anxiety about having to keep all the session caches fresh and end up typing silly things into some of them every 30-45 min so they don't expire or auto-compact.
1）claude（pro）、ChatGPT（business）、Kimi（pro）；各一个订阅； 2）是的；Deepseek（API）； 3）默认的CLI； 4）否； 5）否。
我是谁：一名做基础医学数据分析的博士研究生。
2x Claude Ultra accounts, 2x ChatGPT/Codex Pro. Couple of local hosted models for quick stuff (text formatting, TTS, etc.)
I'm sticking with Default CLI/Desktop apps + built my own to combine them for my specific workflows. Used previously OpenClaw and Hermes but it seems to fall apart when you give it any real work. Ending up giving up because I just don't have to tinker / fine-tune it.
I will say, I enjoy Muse a lot for some reason for personal things. I don't overshare anything I wouldn't mind saying in public.
Within the last month, I basically use Claude Design exclusively for front end work and Codex for backend work. Codex just works but sucks as design. Claude is great at design (with enough references) but can't seem to actually get anything to completion.
I want to make a joke about that last line so badly.
I've had the same experience as you with Claude, and I have to choose a single subscription, so I'm on Codex. I definitely wish I had access to Opus 5.5 right now.
People using DeepSeek - do you have a zero data retention provider based outside China?
Yeah, I use Openrouter with the zdr parameter.
pi, a few custom extensions for pi, rex, gpt + ollama-cloud for a little ds4.1f
as of late i started running out of 5h quota on gpt plus and felt coerced to upgrade to pro. i'm pretty grumpy about it.
I've got a somewhat sophisticated self made setup:
1a) OpenAI API subscription, integrated into an office suite + media pipeline I wrote 1b) OpenAI codex $100 mo plan
2) Current count, 318 additional generative models across LLMs and then media generation of image, voice, music and video. The LLMs are mostly Qwen variants, and the media models are pretty much all the popular media models. These are organized, managed, and API operated locally through WanGP (open source model host) and two additional apps I wrote. One of those apps is a mobile phone pairing remote live video processor, so one can use their phone like a remote live camera at events, and the other is a plugin to WanGP so my software can securely operate WanGP remotely.
3) I use OpenAI Codex in a terminal as a coding assistant. Through that same account I also run the OpenAI Desktop app, and that integrates with my web apps through WebMCP and two additional types of agents in my apps: 1) all my UIs have knowledge assistants that explain how to use the UI they are associated, and they know more complex workflows that UI is a component. 2) My apps have "method actor agents" that are conversationally programmable experts located on specific interfaces, which users then tune to be experts in their work tasks. That's like a smart assembly line worker waiting to be told what they are making. Then, through WebMCP, external agents will interact with my app resident agents, and together they act. Which is kind of amazing. So, in a way, they create a harness through their collaboration, which also includes the user, collaboratively.
4 & 5) no.
My work interest is in creating an environment that promotes deep intellectual flow. One could use my setup and automate a bunch of things and sit back while their brain rots. Or one can also participate in the work at an interactive creative stream level, where the agents and software are doing work but all the creative and impactful decisions delegate to you. You're in complete control over a swarm of virtual expert personalities that work for you, and you make their decisions. The work is office and media, which is broad.
I’ve moved from paid subscriptions on Codex to Claude to Muse.
I now use Muse as my primary agent and then spin up and manage Codex and Claude agents as needed on an Oracle server.
1) the cheapest Claude sub, $20usd I think.
2) pretty much just opus
3) claude code - but with Orca + app (being able to dev remotely via the app has been great for personal projects)
Work has me stuck with Cursor, which I don't find as good, even when using claude models.
1. Claude pro monthly
2. Kimi on OpenRouter, Abliterated.ai as needed
3. Claude Code (with all models), Claude Desktop, tried pi.dev, opencode
4. No
5. No
Main use case(s):
- Software applications
- One-off automations (calendar sync from scraped data, email alerts)
- Deep research and planning for personal stuff
Wow I am pretty sure I read thru all comments and mignt be the first to say I use qwen 3.8 advanced because it's free
1. Claude Ultra for personal, Claude Max for work, Codex $100/mo
2. OpenRouter, DeepSeek, Luna, and testing out models here and there
3. Claude Code and Codex for the respective subscriptions. Played with OpenCode and Pi some as well
4. I've tried a number of these tools and settled on my own tool (not released) to oversee my agents
5. Played with OpenClaw and Hermes in the past, cool ideas but execution fell flat. I currently use Dots and Duet for little things (mostly monitoring communications). Dots is much better than Duet today but I hope Duet gets better.
gptel + gptel-agent in emacs, hooked up to a Claude API token and ChatGPT oauth provided by work, plus some custom gptel additions (tools and tool invocation inspection mainly).
Personal exe.dev subscription.
Muse and cursor. It beats every custom setup I had by far on ease of use. The cost is trivial.
At work, we're using Kiro. I've got no complaints.
In my personal life, I'm extensively using Hermes. If you look at some previous comments you can see what I use it for. I genuinely think personal AI agents are the future, the way I feel when I use it is how I felt when I tried out the first iPhone. In 5 years time, I think most people will have a personal AI agent in the same way the most people now have an e-mail address.
1) Claude Code (Claude Teams Premium seat) plus pi.dev (self-hosted GLM-5.3) at work, Claude Pro at home. Opus 5.5 Extra effort for ~99% of the prompts.
2) Not really, messed around with Openrouter and Deepseek for some stuff that Claude refused to do
3) I use Claude Desktop (Linux) and a heavily modified pi.dev at work VSCode as IDE. At home just use the iOS app and Claude Desktop (Windows).
4) No
5) I have Hermes Agent setup at work and home but don't really use it.
1. Claude Max - two (personal and work), ChatGPT plus, OpenAI (PAYG at work), Gemini PAYG (mostly for automated transcription). I could probably cut my personal Claude Max account, but hang onto it for the constant stream of personal project ideations.
2. Yes, per above
3. Mostly Claude Code, occasionally OpenCode and Codex. Rex/Ghostty as the interface to it all.
4. Yes, lightly, a custom orchestrator
5. Not actively, have tried most ex Grok Bot, but haven't taken the time
For now, I have decided to refrain from using LLMs because I consider the technology a very serious threat to humanity. So my answers are:
1) None
2) No
3) None
4) No
5) No
At work Claude Desktop with Opus 5.5. At home Deepseek Harness, 5-15USD per month. I don't like CLI harnesses. coz it looks too monotonous. I prefer default agent orchestration of the harness, since its their job to take care of it. not something I should spend my time and efforts on. If I did, high chance it will get out-dated soon. I do not trust agent platforms. Lethal Trifecta comes to mind.
1) None. duck.ai for random questions/research for private use. No personal coding project or anything like that.
2) Yes, Anthropic models usually, at work. Pay as you go API. Company is trialing limits, routers and open weight and even local models. Since the Anthropic 5.5 models none seem reasonable to use any longer (for now at least)
3) Claude Code at work. Tried Aider early on and it failed to even read its initial context (Claude models like a year+ ago). I dropped it dead after that - unusable if you need me to specifically "load" and point our files. That's not agentic coding, sorry. Have not used the actual features my IntelliJ Ultimate subscription provides in quite a long time, though I loathe to tell IT/finance that. I do use it from time to time to double check what Claude tells me. Used `pi` w/ local models (some gemma and qwen that'd run on my hardware. nothing usable w/ the work code base - or much anything for that matter, like I'd not even get through an interview coding challenge with that if you ask me)
Did use Codex maybe like 8 months or so ago to ping pong implementation plans and reviews between GPT and Anthropic models but then just stuck with Anthropic for some reason.
4) Nope.
5) Nope. But we do have our own in-house review agent and software factory agent. Like think CodeRabbit et. al.
1) GPT plus, z.ai GLM Coding Lite-Monthly Plan ( legacy ), Opencode Go
2) personally, no. At work there is a Gemini Enterprise and Kiro Pro+ subscriptions
3) pi.dev inside herdr; both at home and work. I just pile the APIkeys and stick to same harness.
4) mostly, pi-subagents has an orchestration mode
5) no.
I use Orca as an “ADE” (mostly for the multi-worktree interface and mobile app).
In my day job I use Claude Code & Codex subscriptions (Codex is mostly for review, second opinion, and as a backup).
Personally I use both oh-my-pi and Hermes with a Codex subscription. I tend to use OMP more now though, since I find its interface a bit nicer when working on the laptop.
Nearly identical setup except I splash some Chinese models in there like DeepSeek v4.1-10-17 and GLM 5.3
After trying so many different setups, now I just use Claude-- the only thing interesting about my setup though is that I separated a "work-claude" and a "personal-claude" install on my laptop, which isn't officially allowed and took a bit of wrangling. But I basically had to do it legally.
1) Claude code (Max) at work. Claude code (Pro) personal account for side projects. I do still have a ChatGPT subscription. I'm tempted to cancel it, but I like their app. Gemini is bundled somehow into my google subscriptions. I use it occasionally, but I'd be just as happy to not be paying for it.
2) I do use Openrouter for side projects that need an embedded AI. For example, on Halloween I made a talking pumpkin to greet trick or treaters as they came to the porch.
3) Mostly code. I've tried muse, grok, gemini, codex, and pi w/ openrouter, but I always come back to code.
4) no
5) I have a grok bot account I use to manage personal stuff. It organizes my email and calendar, keeps me up-to-date on news, does health tracking, trip planning, and shopping research. As an example, there's a website where I order lunch for my kids. I like to know each morning if I need to pack their lunch, so I have grok bot log into the website and sync it to my calendar. I then get a notification in the morning with everything I need to start the day, like whether to pack lunches, what the weather will be, or if I need to help my kids study for a test.
Tried just about everything over the last couple of years. Settled on the unified Cursor, Grok, Grok Bot plan for about $200 / month. Doing everything within herdr for programming and Grok Bot for all non-programming related AI life stuff.
It's a solid groove. Grok Bot helped me get so much stuff off of my plate.
$100 ChatGPT sub $100 Claude sub
Most coding via Opus 5.5.
Codex for text and some non-coding tasks.
Separate virtual desktop and clone on dev drive with VScode for each task (much prefer its diff tool and use it to launch builds etc).
I never come close to running out of tokens because I still do code reviews often enough to throttle usage.
Herdr, Pi, OpenRouter. I haven't found that I've needed much more than this for quite some time now.
I'm playing around with some ideas on top of pi durable. Cascading contexts exposed over http. Mostly deepseek, glm and kimi in there. https://github.com/resolveworks/key
1) What AI subscriptions are you personally paying for? (Claude, ChatGPT/Codex, Cursor, etc)
No. I moved off of subscriptions earlier this year. I don't trust the subscription model from any frontier lab. Just look at what Anthropic has done. The first time I tripped a safety classifier on a seemingly irrelevant task, I canceled everything.
2) Are you using models outside (or instead) of those subscriptions? For example, pay-as-you-go APIs, OpenRouter, or locally hosted models.
Local first. I have a box running Muse Spark on my desk. My main inference box is a launch day DGX Spark (I saw the hardware shortage coming) and has my entire local AI stack. OpenWebUI and everything that feeds it. I moved away from SearXNG and started using SlopSearX (it's quite good).
Everything else goes to Openrouter. I use few models but GPT-6 Luna is my go-to, sometimes I switch it up. Really depends on the task I need to perform.
3) What coding harnesses and interfaces are you using? Are you sticking with the default CLIs or IDE integrations, or using alternatives like OpenCode, Pi, Aider, Muse Code, or custom-built harnesses?
Github Copilot for work. It's meh. Personal projects and general personal use? I started with OpenCode and have since migrated to Pi (omp specifically).
4) Are you using any agent orchestration or swarm tools?
I thought about it. I haven't really felt like I would derive value from changing my workflow and tools. It took me a while to find a good mix of utility, control, and efficiency and haven't wanted to change it until at least next year.
5) Are you using broader agent platforms like OpenClaw / Hermes Agent / Grok bot / Dots / Muse?
I have a dedicated Mac Mini to run Hermes. It is pretty much integrated into my entire life now. I could live without it, but automating general purpose things is the entire point of why I put that machine together.
How is this different from this time last year? Radically different to say the least. The pace of change has been EXTRAORDINARY to say the least.
Claude Code Ultra as the primary agent, Codex Pro as a second-opinion reviewer, with Rex as my terminal and multiplexer.
1) None
2) Openrouter. I don't have a preference right now. All the top open-weight models work fine for my projects. Currently using Mistral.
3) I used pi for a while and lately I started using openrouter's ori coding agent. Working great so far.
4) None
5) None. I tried Muse for a while and it was too slow to do anything useful.
1. Claude Ultra + a Codex Plus. Probably going to drop Codex soon, GLM is gradually taking on what I intended it for.
2. GLM-5.3 and OpenCode for cybersecurity
3. Primarily custom built harness. Rust centric, heavy on the production engineering gates and a bunch of graph networks for context control.
4. Custom.
5. No.
My general ethos is that if taste is what is important going forward, my code generation must have it's own taste. If you can download my tooling off the shelf, I am replaceable by anyone who can do the same. My end goal is to drop Claude Code once the harness prices start to drift to API pricing. In the meantime, anything that can be automated into traditional and efficient code will be one more thing in my control regardless of future token prices. When I hit my limits for the week, I spend the remainder of the waiting time going over the framework data to see what are my largest sources of consumption and iterate in automating that more.
Codex on Pro 100, Claude on Pro 20.
I don't even understand how people can spend thousands of dollars a month on LLM tokens since I've accomplished multiple major projects with this setup. In general, Claude's better at frontend and debug and Codex is better at feature implementation and research from what I've found, which is why I have both.
2. No. GPT 6.1 Sol is very token efficient that I pretty much never run out before the end of the week anyways.
3. Default Codex/Claude app. I only use VS/VSCode to hand-edit/git sync things when needed.
4. No. Swarm is a waste of token imo, and a single good human in the loop beats any automated orchestration tool.
5. I find it unnecessary.
1) Claude Ultra, Codex Plus, SuperGrok, Gemini Pro.
2) No.
3) Opus 5.5 on high for all the work. Reviewed by Codex by default. Additionally by Grok and Gemini (agy) for critical work.
I switched from Claude CLI to a custom-built harness because I got overwhelmed by walls of text in parallel sessions. Now I have a canvas per project where I spatially lay out cards. Each card spawns an agent in a worktree. The chat with it is condensed to WhatsApp style chat bubbles with all the thinking and tool calls folded away. The result of a card is a demo, usually a narrated video that I can comment on to trigger another iteration. Everything voice first, I mostly speak. Phone companion so I can keep working on the go.
I used to work with up to 6 Claude CLIs in parallel and couldn't go further. With my custom UX, I can do 20+ with fun. Now the bottleneck is actually my machine, so I added remote worker support to my harness to distribute the load to more machines.
(I open sourced the harness in the hope of sharing the fun. If you want to try it: https://github.com/sadilek/obeya)
4) No Paperclip or Munder Difflin or anything. I actually think it doesn't make sense to create persistent agent personas. Just spawn an agent when needed, give each full access to all skills and docs, and they'll figure it out.
5) Hermes Agent running, sending me regular news research reports, but frankly, it's just a cronjob and the whole Hermes setup isn't really adding value for me.
1. Two Claude Ultra, ChatGPT plus, Gemini through credits
2. Pay-per-token for work through Vercel AI Gateway (mainly same models + grok, but trying out others too)
3. Claude Code, Codex & OpenCode
4. Homebuilt https://svall.dev/
5. No
Google Antigravity CLI using Gemini Flash 3.8 most of the time. Amazing what it can do.
VSCode + OpenRouter + flavor of the day cheap model.
I spend about $5 a month with this setup.
1. Claude Ultra, ChatGPT Pro 200, SuperGrok, Kimi Pro, Cursor Pro, Google AI Pro, Muse Code High Usage. It comes out to ~$500/mo, mostly from Claude+ChatGPT
2. I use OpenAI/Anthropic APIs for various workloads (app features, automations, etc.), but not for personal inference/coding. Recently both of these providers started giving $200/mo in API credits, and OpenAI already gives free tokens, so I don't usually pay for this usage.
3. I always use the provider's preferred CLI: Claude Code, Codex CLI, Grok CLI, Kimi CLI, Antigravity CLI, Muse CLI. I use this with herdr to manage everything, and then a small custom app to keep track of my usages since I hit limits quite often.
4. No. Just the CLIs. For very large projects I've found Claude Code to be the best -- it can very easily manage many subagents.
5. I've tried OpenClaw, Hermes, Dots, and Muse. I really just haven't found great use cases for these tools
I use Claude & Codex for most of my personal coding/projects. For the other providers I mostly want to have broad experience with different models/harnesses. It's very interesting to see how each performs and what kind of tasks they do/don't excel out. The only con of having cheap plans is you really can only give them simple tasks.
---
At work I use AWS Bedrock w/ Anthropic + OpenAI models at API pricing. Each employee is given a rolling 28-day budget w/o access to Fable/Astra. I use the same Claude Code/Codex CLI + Herdr.
For personal stuff, a $10/month OpenCode Go subscription with OpenCode, mostly running DeepSeek Flash. I've yet to run out of tokens.
Work provides an enterprise Claude subscription for work stuff, I've never hit any limits on that either.
https://github.com/mkornreich/llm-desktop-electron with OpenRouter. I love Claude but Anthropic is expensive and too restrictive
Claude Code. Everything else default!!!
We built our own harness local & cloud models, test it out: www.vekmem.com
1) OpenCode Go $10, though I'm just experimenting with it. Before I have just been using the free models available in OpenCode, since sometimes its hard to justify a subscription when you're unemployed.
2) I tried a local Qwen model but my poor old 1660 Ti did not enjoy it, and it wasn't all that useful.
3) I primarily use OpenCode running in a podman container for actually working on stuff. Also use the default free chats with Gemini 3.5 Flash for general searching, ChatGPT for brainstorming, and Claude to answer more code related things if they come up.
4) Nope.
5) Not at the moment. I've been eyeing using Hermes, but since I don't feel comfortable paying much at the moment I haven't really tried.
Only GitHub Copilorpt for documentation. Sponsored so cost nothing.
1. No
2. OpenRouter PAYG - glm or opus for most things
3. OMP or CC
4. Zellij/Vim — some things never change
5. No
I'm using Claude Code (Ultra) across a multitude of tmux sessions, over two or three laptops – largely so long running actions don't happen on the actual laptop I take with me (nothing worse than waiting around for an agent to finish before you can leave a café). To make it a little easier to visualise across those sessions I built https://belfry.robgough.net which is like herdr in some ways, but just uses standard tmux.
I also gave in after seeing my co-founder use it, and gave GrokBot a go. I use this for everything that's not coding or design... so marketing, email reading, right now it's currently inputting my meter readings to my electricity company after emailing the building management company and getting photos back. It's super useful at little chores that eat up time. I have it daily check the status of my various projects and services. I love it tbh. I've also built my own version, which I might move to... though now it's all set up it's quite sticky, and it already spotted an error my accountant made on some taxes so it's already saving me money and I'm feeling quite positive about it.
I'm getting a bit frustrated with having to run things on my laptops though -- it feels a little pets not cattle, and wish it was all in the cloud. I'm toying with building out an agent factory, but frankly there's no super-cheap way of doing this so I'm holding off, though I can see me building a tool that connects those laptops to it creating sandboxes with Apple's `container` (which I'm using successfully on my grokbot replacement).
have you tried https://www.droneship.run
Mainly CC Max with some API calls to OpenAI for different things.
No harnesses right now. Ok occasion using models in LocalLM.
Haven't got into agent orchestration much other than Paperclip earlier this year.
None. These things aren’t that good.
1) 20$ Claude subscription. I used to have 2 20$ subscriptions but not anymore.
2) My main model is Deepseek flash via the API.
3) Claude Code CLI inside OpenShell. I like Claude Code's UX. Lots of custom skills and scripts.
4) No
5) No
1/2) OpenAI Business Premium subscription and a locally hosted Qwen 3.8 Flash Next running on a DGX Spark.
3) I use both Opencode and Pi.
$200 ChatGPT + $200 Claude, I don't pay for it, agents quant trade and pay for their own subscriptions. I then use the rest of the tokens to build stuff.
Im curious how hardware/firmware engineers are using Claude, et. al. for work. Tools/MCPs that are worth it?
DeepSeek Flash with kicad + autorouting churned out a couple of working PCBs with minimal prompting, and customized ZMK itself for some custom input devices. At least for casual stuff this is a solved problem
1) Xiaomi Mimo Pro 2) I purchase periodic credits through Tailscale Aperture 3) MiMocode for working with Mimo, Pi with the others 4/5) Nope
I use Codex and Claude Code with piles of project-specific skills and AGENTS.md files.
For my main work at my startup, we're very careful with how these are setup.
Everywhere else, I simmer in chaos.
I also have strong boundaries around generated prose. I don't like making other people read AI slop, but I don't mind having AIs read it.
My own harness (nothing intricate, but to my taste), DSv4 1017 on my own hardware
Keeping it simple.
1. Using ChatGPT with a Business Premium seat (not personally paying for it).
2. Also have some API keys for some open weight models in Europe, but currently not using those.
3. I've personally used OpenCode, Claude Code and Codex and of those, I liked Codex the most, till bugs kept on accumulating and they started doing all kinds of shit like that daemon recently. I now switched to https://github.com/capocasa/3code which is a small, simple agent harness but with built-in sandbox, unlike so many other simple harnesses. I really like this one, plus the author is very responsive to feedback
4. No. And I don't think that stuff works in the long run with current quality of code output
5. No.
I've made a multiplayer environment for me and a few friends, where we can prompt (currently three) emergents and experiment with various models and instructions for them.
I've rarely been so inspired by a project.
It's publicly readable, but writable by a group of ~20.
https://magenta.cryptograss.live/
Codex (Pro x 1), no outside models, custom harness, custom orchestration, custom agent harness.
I avoided building anything custom until I had to, and I avoided adding more than a single agent until I had to.
My preferred way to use AI for coding has been to work closely with a single coding agent with the code open. That changed wth 5.6-Sol and a handful of skill files. I stopped needing to watch closely to keep it from fucking up the codebase. With the right request, that combination will get the job done 9/10 times.
That unlocked a new workflow: as soon as a change is requested, a coding agent reads it and gives me a PR. I want a complete proposed code change in front of my eyeballs by the time I open the change request myself.
So I have a custom solution in place to manage a handful of agents aligned with this workflow.
Custom solutions are so cheap and easy now... it's really hard to put up with solutions that are only mostly right for the job.
I still just pay $100/mo for Claude. Use the CC extension in VSCode. I’m actively investigating various ways of getting a Hermes Agent going on my homelab with either a local model, Openrouter ZDR models, or both but I’m unlikely to pull the trigger on that yet. I would never use Muse, Grok Bot, etc as a personal agent.
1. Claude Ultra, three accounts. Mostly to get enough Fable usage.
2. Early in experiments with locally hosted GLM Next, not writing code with it beyond experiments.
3. Claude Code, system prompt, minimal skills; mostly bash scripts for common actions.
4. tmux send-keys for messaging between panes.
5. No. Claude Code does that, too.
Isn't Opus 5.5 better than Fable 5.1?
On some benchmarks, yes. For real software development, no.
I use OC with various subs, and am testing DeepSeek Harness, JetBrains Thinkrail, and Volt. Extending the harness with plugins is my main interest these days.
I mostly work on problems at the intersection of manufacturing, mechatronic design, electronics, software and physics. So I'm not really building software per-se which likely accounts for the probably slightly weird choices here.
1. Claude Max (upsold me as of today); generally Sonnet 5.5 with the occasional Fable 5.1 when it gets stuck or I am executing a high value step; and Perplexity Pro (legacy).
2. Mostly local voice-to-text w/Parakeet Unified EN 0.6B via Handy[0] (awesome; use it!). Not for coding, but for note taking, AI interfacing, and general typing reduction and RSI avoidance.
3. None. I tried Pi and after much kerfuffle I don't like any of the harness interfaces. I find the type of work I do they are not very useful, because I often switch between variant approaches, planning, data, and don't want unrestrained context bloat.
4. No.
5. No.
I would add that I am very interested in obtaining good enough hardware to run this level of model locally at any pace, believe Qwen is basically there but too slow locally to use as a daily driver on my current rig, and can see myself making that investment in single digit years. Not right now, the hardware is overpriced for what it is versus venture funded cloud AI subsidies.
[0] https://handy.computer/
Dont use it at all. Seems like a waste of time for what I do as far as I can tell.
none; no; none; no; not needed; no; none; unnecessary; no thanks; and not interested. "I've seen enough, kiddos."
- LLM reverse proxy
- agents work in their own lxc spawn automatically in a proxmox node, where they can go wild without sandboxing or isolation, network is firewalled too
- all sync sessions and everything to a centralized agentsview in a realtime, where I get to track and watch it from anywhere on the go
- i just ssh into the lxc if i needed anything
- different harnesses depending on the task, maki is great one, but I use others
- I also have other stuff but I won’t mention it here regarding the tokens for obvious reasons
I tried most of the tools like herdr or paseo or desktop clients but they never clicked with me
i dont have one
For personal work. I use Orca.
1. Claude Code with $20 Claude subscription. This used to be a terrible value but now it's a great value, probably because of the IPO. I do a lot of coding and other things and I rarely get close to my weekly limit. I mostly use Opus 5.5 medium. Codex with gpt used to be good value but recently I canceled it cause I kept reaching my limits.
2. When I want a "different pair of eyes" to review things I'm working on with Opus, I use Pi with other model providers and just pay API prices.
3. Deepseek harness with a deepseek API key. Really great value.
4. I have a Muse $5/month subscription that I keep as a backup although I don't really need it.
I also use cloud agents to setup reoccurring jobs. For expensive agent runs, like visit 5000 websites and extract something from each, I would set up something cheap with an agents sdk and for the model use either Muse, Deepseek, or GPT-6-luna using their batch API.
I maintain all my dotfiles and agent files / context in a github repo so that I can switch easily between harnesses and providers.
Claude code is all you need
Yes, given that you can route the other models into it this is sadly true.
1.) Claude or ChatGPT Chatbot.
2.) Text Editor or IDE.
3.) A compiler/interpreter/SDK and potentially other dependencies.
Super lame and boring, but that's basically it.
As for my day job, it's the same minus #1 -- AI is banned.
1. Claude Max
2. No (only the API for projects)
3. Buzzpot.app (my app)
4. No
5. No (I tried them, but couldn't find a use case for them)
Opencode with $10 Opencode Go plan until his heinous Clammy Sammy lets me have enough Ram for a local inference setup.
I run
2) kapitsa + pi + Qwen 3.6 35B a3B q4. pi is the harness, kapitsa is the inference engine
people that I considered intelligent have associated Qwen with "chinese cloud" when it implies not using a cloud at all. I think that's weird. I think there is a "western" centric view that limits perspective to using A cloud, instead of aspiring for locally run
1) I also pay for Claude, I should downgrade again to the $100 plan, thanks for the reminder, but currently the $200 max x5 plan
4) no. I handroll my agents
5) I use Claude desktop as an agent platform
here is my production grade setup for agentic ai on windows 11
- setup windows 11 with Rufus boot to disable online MS account creation https://rufus.ie/en/
- [download window 11 x64 iso from here ](https://www.microsoft.com/en-us/software-download/windows11)
- then download vmware workstation pro (you ll need to create a free account to get the download link on their website)
- install vmware workstation pro
- turn internet off
- plug your pendrive in with the rufus w11 x64 iso on it (mind ya, this pen drive ll first need full formatting before burning a w11 image to it)
- press f2 to reach bios in vmware settings and make it boot from pendrive
- turn internet off on your machine (DO NOT keep it on)
- install w11 (keep a 100gb for vmware)
- without updates it ll load real fast and you wont need an MS account (rufus takes care of it)
- while it installs, download revo uninstaller pro on a separate machine https://onefiles.net/9d170d01078634b5 (latest version 5.4.x) and put it inside a second pendrive / portable hdd
- once w11 loads, it ll ask you to enter a password, keep existing password blank and enter literally anything for a new password
- voila you are done with w11 setup inside vmware
- now connect the 2nd pendrive with revo and install it (need to clean uninstall many tools, hence we need this one)
- now go to search bar and type check for updates and update everything on w11 inside vmware
- next stop get unigetui from ms app store or from here https://devolutions.net/unigetui/ (it is a package manager responsible for installing just about everything else with auto updation of all packages)
- the open unigetui and search for the following packages vlc, winrar, winhance
- winhance is the fully customizable debloater you ll need to turn off all telemetry (we ll debloat after installing all the other requirements)
- next go to ms app store and install python or open powershell and type python
- the open unigetui and search for the following packages vlc, winrar, winhance
- winhance is the fully customizable debloater you ll need to turn off all telemetry (we ll debloat after installing all the other requirements)
- next go to ms app store and install python or open powershell and type python
- mind you this installs a python version manager that lets you install and uninstall multiple python versions
- after this python version manager is setup, install a specific version of python like
``` py install 3.13 ``` - next we need a node.js version manager, use fnm for it
https://github.com/Schniz/fnm - once installed, it may need to be added to environment variables manually
- once again choose a specific version of node.js and install it using
``` fnm install 23.12.0 ``` check fnm --help for commands - next we need a java version manager, use jvms for it
https://github.com/ystyle/jvms
- pretty self explanatory
- now install git and again it ll give you some trouble over environment variables, command not recognized etc, fixable with some gpt help
- at this point, we got git, node version manager, java version manager and python version manager and are ready to rumble
- install vscode/ visual studio whatever
- install codex
- install claude
- install opencode (both cli via npx and desktop app) and check their plans out https://opencode.ai/console/go
- install ghidra , ghidra-mcp (mind you open ghidra-mcp first because it only works with specific versions of ghidra and java) this is the reason we installed a java version manager
- check out setup for rea https://github.com/morluto/rea requires npx setup for rea and ghidra both, read documentation
- run winhance, check recommended settings under optimize and every other tab and restart the vm a few times for full debloat with 0 telemetry services
- congrats you have a full blown agentic environment at this point (if I missed anything i ll come back and update)
Ok I don't understand how people can deal with the information coming from a fleet of agents burning 100s-1000s of €/$ as month.
I have a Claude Code pro subscription, and mostly have Opus as a thinking buddy, we have thoroughly designed a data model together, we have implemented it in Django. Then we designed the RBAC system and now we are making views per role.
It is so intense! Constant decisions, walls of text with pros and cons to decisions. I mean the speed is amazing, the code quality is also amazing, but the amount of information I get to process is just insane, and I enjoy it less than the flow-state writing of code. I think I'll get used to it, and the speed is motivating, but man. I can not picture keeping 10's of agents on tasks. What am I doing wrong? Should I trust the system more? Or am I just mostly working on decisions now and the 1000$ token burners are not bothered with that because they have a full backlog, where I need to fill it myself?
I deal with the overload by making abstractions (a visual of the data model/db schema with color coded relations etc, a table for RBAC, think about permissions in "Axes"... I'm always looking for smart tricks to "think better", to be able to hold more information and reason about it. What must exist in me, what can I outsource?
Btw, as a FOSS person I so hate that I am not owning my means of production anymore, I have a home battery for all my infra, very proud of what I build, but I can't run my agents :( I really hope that will change soon (for privacy and for control reasons.) I pisses me off that I can't get nice hardware, meanwhile I can get a Muse or Dot, which is an LLM with a VM with up to 12 GBs of RAM, for free, paid for by their borrowed money. This whole idea makes me die a bit inside.
For the company I do all IT for we have a Mistral Team subscription, I find it hard to work with anything other than Claude Opus, I really force myself to do it but I often think: Pff, Opus would have gotten this right. But I remember feeling the same when I switched from Google to DDG (3000 years ago) and I never switched back. So it's probably just familiarity.
I also have HermesAgent on a nuc here in front of me, it makes small automations: Scans for positions or projects for my company, parses RSS feeds for relevant cybersecurity news, researches things when I ask, stores notes and articles with I can interrogate as a "body of work" <- love that btw, it's like an external memory where the faintest glimmer of a memory can help you dig up that one nugget of knowledge. Or you can find trends in TFAs you love. You can really get to know yourself a bit better. Most things it makes are LLMs out of the loop, just automations build from simple Unix-style tools that I could have made myself but never took the time for. It's also nice that it solves edge cases itself. It burns less then €3 a month on deepseek or so (just processing public data for me). I keep it on a separate vlan and away from important stuff.
I don't think you are doing anything wrong; it is overwhelming. But I also don't think "trusting" the system more would help. IMO, now coding is a bit more like maintaining critical infrastructure; it may be built quickly, but you need to keep an eye out for random decisions agents are taking that will cost you (money, security...) in the future. I have seen Claude making many sneaky decisions and (maybe) coincidentally burning more tokens by doing so
You use Claude as a "thinking buddy", meanwhile those who spend $500+ on LLMs have shut their brains off, and have delegated all thinking to the LLM, producing tons of slop.
herdr, opus 5.5 (max), codex (max), and I am still refining how they all communicate, currently codex became unwakable at some point while claude can be woken with a message just fine.
devin
1) anthropic ultra; single sub
2) payg on opencode
3) standard claude code and opencode driven by t3 code
4) nope
5) grok bot - at work, to manage my tasks, and to highlight what happened on slack when i was off; random one offs there
I built a "software factory" on top of pi agents so I can automate most product dev work
moved to openclaw for everything, with occasional claude usage for tricky bits. I have $200 of both codex and claude subs.
Openclaw is the answer to question "do X if Y happens". An agent per each type of Y channel manages the flow and memory. And user input is just another case of 'Y'. It is a very powerful model I am having lots of fun with it.
the plumbing is a chore though, thinking of trying out grokbots to see if they solved it.
Right now, it's Claude Code on Opus 5.5 w/ herdr to manage multiple sessions at a time.
1.a) At work: I can choose between Cursor and Claude on a meter. Been using Cursor for a long time and happy with it until they started to force Grok down our throats. I was happy with Composer 2. Recently switched to Claude Code. The company also runs some open weight models in-house and I have been using them through Claude Code itself, with VSCodium as my Code Viewer. I almost never need anything stronger than Sonnet at work - 80% of which is YAML and Terraform these days (or rather markdown, if you really think about it).
1.b) For personal stuff I maintain an Opencode GO subscription, USD40 tier. I also have a Google One subscription which is supposed to have some AI capabilities, haven't tested it yet.
2) Besides my Opencode GO subscription, I sometimes add a few bucks to use Zen. It's rare: my USD40 is usually fine for the month. I've recently started using Nvivia NIM free tier models for some time insensitive tasks too. I have a Qwen model available on my Mac (M1 PRO 64GB) but haven't found it to be useful, as I need my laptop for everything else I'm doing beyond inference.
GLM 5.3 Flash and Deepseek 4.1 Flash do most of the work for my personal stuff. Nemotron does the time insensitive stuff, which is mostly summarizing text. I've been using GPT Luna 6 a lot because it's currently cheap as hell, so I'll often talk to GPT Luna and hand off coding to GLM or Deepseek. I find that Luna communicates MUCH better than the Chinese models - I assume because it does not go through the extra language barrier step between English and Chinese. GLM, Qwen and Deepseek are incredible tool callers, but their prose is insufferable.
3.a) At work I'm using Claude Code for the Anthropic models and Opencode for the self-hosted ones. Ditched Cursor recently. I've been meaning to try to integrate the self-hosted models into Claude Code but haven't yet, so hand-offs between the two harnesses involve a lot of markdown.
I've switched out of Cursor because A. they've been bullying us into using grok for everything and B. The company setup would not support the use case of mixing the self-hosted and closed models in the same harness - this limitation is on our side.
3.b) For my personal stuff I've been using mostly Hermes. Some actual coding goes into Opencode, often piloted by Hermes itself. I'm using Zed as my code editor, but with the AI features disabled. My opencode subscription is actually configured in Zed, but I prefer the UI without them.
4) No. I don't really need them. I run 5-6 parallel sessions at most, otherwise I can't keep track of what's being done anyway. Each parallel session is on a different task, so I generally don't need them to spit enormous amounts of code. I rarely have to wait more than a few minutes for a task to finish, and I'm always thinking or writing or managing something else in between anyway - not looking at the screen waiting.
5) Hermes. Generally happy with it - despite it being extremely unstable between updates, with frequent regressions, weird UI bugs and half assed features (bot mode, anyone?). I have 3 distinct instances (laptop, home server, VPS) doing different things - all hooked up to the same memory provider (honcho) under a single user profile, and a handful AI profiles on each. All have limited access to my Obsidian Vault, which is synced between them using Obsidian Sync.
I generally use the laptop instance for day-to-day use: planning work, assigning tasks to myself and to the other agents, capturing ideas, light coding, etc. The homeserver instance manages a limited set of features on the server itself, from upgrades, documentation and more devopsy stuff to organizing files, managing transcoding queues, etc. The cloud hosted version automates is not very used yet. I plan to have it pilot an n8n instance I keep to automate some boring stuff, like filing weekly tax records and automatically forwarding bureaucratic e-mails, nothing fancy.
> 5) Hermes. Generally happy with it - despite it being extremely unstable between updates, with frequent regressions, weird UI bugs and half assed features (bot mode, anyone?). I have 3 distinct instances (laptop, home server, VPS) doing different things - all hooked up to the same memory provider (honcho) under a single user profile, and a handful AI profiles on each. All have limited access to my Obsidian Vault, which is synced between them using Obsidian Sync.
I am just getting into setting up Hermes on a VPS. If you have any high ROI tips/tricks/advice on how to make this work well, please let me know!
Appreciate it!