I've been working on this clone of mistral vibe, the idea is to have a cli that you can trust is not calling back home for any reason, I have done as much work as possible to remove any trace of phone back anywhere, hence the pri prefix to the name privibe.
It's been tested 99% of the time against a local llama.cpp, I have also used it against open router with no issues.
Then on top of it, I want to be something private that you can use locally, I've been also working on this branch of llama.cpp https://github.com/alainnothere/llama.cpp/tree/disk-cache-ev... that I use along with privibe that stores conversations to disk, so if you change conversations, or have multiple windows open the conversation goes to disk and you can resume from disk at any time instead of having to reprocess everything again, here https://github.com/alainnothere/privibe/blob/main/docs/resum... you can see the server being killed and then continue the conversation with minimal downtime (as opposed as waiting 5 mins to reprocess everything)
I have also try to make the thing be as cache-efficient as possible, and privibe has a lot of work to avoid modifying for any reason the conversation, so you can always resume from disk and for no reason things change from whatever is stored on disk
Other parts of work:
- llama.cpp modified the chat template to avoid the issue with cache not matching that causes reprocessing
- privibe has on top of the read and write tools "hashed" versions that use line and checksums to make file editing precise, along with fixes to correct issues like the model duplicating lines and messing with spaces (as long as the changes are "best effort" correctables)
- did I mention it does not call phone anywhere?
- If you use the combo of privibe + llama.cpp branch reprocessing conversations should not happen (unless there is that scenario I'm not seeing)
- Both have been tested extensively, and works correctly in linux and windows git bash (I know it also works on powershell but I have not tested it as extensively as git bash, because I'm lazy and I have to set up a lot of stuff... I have tested it for months, I've been keeping the branch of llama.cpp up to date with the upstream until I hope not, something is changed so drastically that I cannot trivially merge back into my branch)
Yes, it's claude assisted, and what I offer on top of the "anyone can ask claude to do that" is the dogfooding/testing of it, for my philosophy check here https://news.ycombinator.com/item?id=49289341
I've been working on this clone of mistral vibe, the idea is to have a cli that you can trust is not calling back home for any reason, I have done as much work as possible to remove any trace of phone back anywhere, hence the pri prefix to the name privibe.
It's been tested 99% of the time against a local llama.cpp, I have also used it against open router with no issues.
Then on top of it, I want to be something private that you can use locally, I've been also working on this branch of llama.cpp https://github.com/alainnothere/llama.cpp/tree/disk-cache-ev... that I use along with privibe that stores conversations to disk, so if you change conversations, or have multiple windows open the conversation goes to disk and you can resume from disk at any time instead of having to reprocess everything again, here https://github.com/alainnothere/privibe/blob/main/docs/resum... you can see the server being killed and then continue the conversation with minimal downtime (as opposed as waiting 5 mins to reprocess everything)
I have also try to make the thing be as cache-efficient as possible, and privibe has a lot of work to avoid modifying for any reason the conversation, so you can always resume from disk and for no reason things change from whatever is stored on disk
Other parts of work: - llama.cpp modified the chat template to avoid the issue with cache not matching that causes reprocessing - privibe has on top of the read and write tools "hashed" versions that use line and checksums to make file editing precise, along with fixes to correct issues like the model duplicating lines and messing with spaces (as long as the changes are "best effort" correctables) - did I mention it does not call phone anywhere? - If you use the combo of privibe + llama.cpp branch reprocessing conversations should not happen (unless there is that scenario I'm not seeing) - Both have been tested extensively, and works correctly in linux and windows git bash (I know it also works on powershell but I have not tested it as extensively as git bash, because I'm lazy and I have to set up a lot of stuff... I have tested it for months, I've been keeping the branch of llama.cpp up to date with the upstream until I hope not, something is changed so drastically that I cannot trivially merge back into my branch)
Yes, it's claude assisted, and what I offer on top of the "anyone can ask claude to do that" is the dogfooding/testing of it, for my philosophy check here https://news.ycombinator.com/item?id=49289341