AetherLang is no longer just a pipeline tool. It now features 10 native Agent Nodes:
PLAN: Generates & executes its own sub-flows on the fly.
CODE_INTERPRETER: Writes & runs sandboxed Python for 100% accurate math.
CRITIQUE: Self-evaluates and runs retry-loops to fix its own errors.
ROUTER: Intelligently branches logic to the right expert.
ENSEMBLE: Mixture-of-Agents (MoA) synthesis running in parallel.
MEMORY: Persistent state across executions (Store/Recall).
TOOL: External API access for live data integration.
LOOP: Iterative execution.
TRANSFORM: Smart data reshaping.
PARALLEL: Concurrent node execution for massive speedups.
To top it off, I packaged the entire architecture into a ClawHub Skill (aetherlang-karpathy-skill). Now, any AI (like Claude Code) can integrate this framework to orchestrate its own multi-agent ensembles and reflection loops.
From the kitchen to building an Agentic OS. The AI revolution isn't just for Big Tech; it's for anyone willing to sit down and build.
Check out the code and try it out:
GitHub: https://github.com/contrario/neurodoc
Web UI: https://neurodoc.app/aether-nexus-omega-dsl
I built AetherLang, an open-source DSL for orchestrating AI workflows. It started as a side project while building a Greek handwriting OCR tool and grew into its own thing.
What it does:
- A simple DSL for defining AI pipelines (guard → plan → rag → llm chains)
- 28 node types (llm, guard, rag, transform, parallel, retry, cache, etc.)
- Visual flow designer with SVG rendering and multiple layout algorithms
- Step-by-step execution debugger with state inspection at each node
- Basic performance profiling (per-node duration, token count, cost estimate)
- AI-assisted flow analysis using GPT-4o
- BYOK (Bring Your Own Key) - uses your OpenAI API key, no account needed
- Bilingual UI (Greek/English)
Example flow:
flow ResearchPipe {
input text query;
node Guard: guard mode="STRICT";
node Planner: plan steps=5;
node RAG: rag depth=3;
node LLM: llm model="gpt-4o", temp=0.2;
Guard -> Planner -> RAG -> LLM;
output text result from LLM;
}
The debugger is the most useful feature — you can step through execution node by node and inspect inputs/outputs at each stage.
Background: I'm a cook from Greece with no CS degree. Built this over ~40 days using AI tools (Claude, ChatGPT) for the coding. I designed the architecture and features but needed AI to write the implementation.
Known limitations:
- BYOK keys stored in-memory (not persisted yet)
- Single server, no horizontal scaling
- Early stage — feedback very welcome
Would love to hear what you think, especially about the DSL syntax and the debugging approach.
AetherLang is no longer just a pipeline tool. It now features 10 native Agent Nodes: PLAN: Generates & executes its own sub-flows on the fly. CODE_INTERPRETER: Writes & runs sandboxed Python for 100% accurate math. CRITIQUE: Self-evaluates and runs retry-loops to fix its own errors. ROUTER: Intelligently branches logic to the right expert. ENSEMBLE: Mixture-of-Agents (MoA) synthesis running in parallel. MEMORY: Persistent state across executions (Store/Recall). TOOL: External API access for live data integration. LOOP: Iterative execution. TRANSFORM: Smart data reshaping. PARALLEL: Concurrent node execution for massive speedups. To top it off, I packaged the entire architecture into a ClawHub Skill (aetherlang-karpathy-skill). Now, any AI (like Claude Code) can integrate this framework to orchestrate its own multi-agent ensembles and reflection loops. From the kitchen to building an Agentic OS. The AI revolution isn't just for Big Tech; it's for anyone willing to sit down and build. Check out the code and try it out: GitHub: https://github.com/contrario/neurodoc Web UI: https://neurodoc.app/aether-nexus-omega-dsl
I built AetherLang, an open-source DSL for orchestrating AI workflows. It started as a side project while building a Greek handwriting OCR tool and grew into its own thing.
What it does:
- A simple DSL for defining AI pipelines (guard → plan → rag → llm chains) - 28 node types (llm, guard, rag, transform, parallel, retry, cache, etc.) - Visual flow designer with SVG rendering and multiple layout algorithms - Step-by-step execution debugger with state inspection at each node - Basic performance profiling (per-node duration, token count, cost estimate) - AI-assisted flow analysis using GPT-4o - BYOK (Bring Your Own Key) - uses your OpenAI API key, no account needed - Bilingual UI (Greek/English)
Example flow:
The debugger is the most useful feature — you can step through execution node by node and inspect inputs/outputs at each stage.Live demo: https://neurodoc.app/aether-nexus-omega-dsl
Background: I'm a cook from Greece with no CS degree. Built this over ~40 days using AI tools (Claude, ChatGPT) for the coding. I designed the architecture and features but needed AI to write the implementation.
Known limitations: - BYOK keys stored in-memory (not persisted yet) - Single server, no horizontal scaling - Early stage — feedback very welcome
Would love to hear what you think, especially about the DSL syntax and the debugging approach.