Both search modes exist on purpose: keyword search for terms you already know, semantic search for a question phrased as a question, and a stack filter on each so a FastAPI answer does not arrive for a React problem. episodic_store is what makes episodic_search worth anything later — it records the task type and whether the attempt succeeded, not just what happened. The coordination pool assumes you are not alone: pool_signal marks a topic claimed or blocked so a second instance can see it before starting. agent_bootstrap is deprecated and points at agent_handshake.
A knowledge and coordination layer for agents: a searchable document store, a vector-backed episodic memory, and a pool that other instances read to see what is already taken.
- Keyword search across the knowledge base, and semantic search over LanceDB vectors with Gemini embeddings, both narrowable with a stack filter such as python or fastapi.
- Documents fetched by name in four kinds — skills, protocols, EUREKA breakthroughs and truth patches — with the skills list filterable by name and by stack compatibility.
- Episodic memory: store a reflection tagged CODE, RESEARCH, DEBUG or REVIEW together with whether it succeeded, and later find similar past experiences by vector similarity.
- A coordination pool where an instance signals completed, blocked, claimed or signaling against a topic and the files it affects, with recent pool activity readable.
- Code files parsed into semantic chunks — functions, classes and methods — for retrieval.
- A YouTube URL queued at high, normal or low priority for transcription and insight extraction.
- Onboarding that takes your model, context window, platform, client, languages, frameworks and project goal and returns a configuration, plus live system, memory and hive status.
Nothing — no account, no key. agent_handshake is the intended first call: it takes your model, context window, platform, languages and frameworks and hands back a configuration matched to them.
One command — pip install midos-mcp