# the_librarian

A local MCP server that recommends the few right skills for a coding agent's current task out of a large skill collection, instead of loading the whole collection into context, and learns from whether each recommendation worked.

the_librarian is a local-first MCP server that sits between a coding agent and a large collection of skills, so the agent stops carrying every installed skill in its context and instead asks for the one it needs. The collection stays as plain markdown on disk; agents never load it directly.

An agent describes what it is trying to do in plain language, and the librarian finds the skills whose meaning matches by semantic search, avoids repeating a recent recommendation, and can optionally let an on-device Apple Intelligence model weigh the finalists. Agents only ever read recommendations and report outcomes back; they never edit skills, weights or rankings, and every stage runs locally with Ollama embeddings and SQLite. It is written in Python.

Website: https://github.com/aka-kika/the-librarian