v0.9.1 · Apache-2.0 · Python
Local-first context for LLM agents
An OpenAI-compatible proxy and MCP server over a persistent SQLite knowledge store. Point your client at it, and each turn it:
- weighs, then retrieves your codebase and documents into the context window
- runs no model inference on the default retrieval path
- stamps every context packet with a know/miss contract
$ pip install cymatix-context $ cymatix ingest ./docs --recursive $ cymatix query "what changed in the retrieval path?"
Receipts: EnterpriseRAG-Bench leaderboard (v0.6.4 submission; current release is 0.9.1)
Retrieve
- FTS5 lexical search widened by tags, synonyms, and co-activation
- Fully algorithmic by default: never waits on a model
Deliver
- Assembles to a hard token budget
- Skips what a session already received, so context stays lean over turns
Calibrate
- Every packet answers: know, with a confidence, or miss, with a reason
- Agents can trust what they got, or refresh it
Surfaces
- Proxy: transparent. An OpenAI client points at
/v1/chat/completionsand context rides in with the request; the know/miss verdict is applied inside the pipeline before anything ships upstream - MCP + CLI: explicit. Claude Code tools and
cymatix queryhand the context back to the caller, know/miss block attached - Agents:
/context/packetreturns the agent-grade form: verified evidence, stale-risk flags, and refresh targets to act on