Instead of grepping the docs or trusting a model's recollection of them, the client gets the passages that actually cover the question. What comes back is quoted from the source, so it can be checked against it.
A server that turns the Gemini CLI's markdown documentation into a searchable index and exposes it as a single MCP tool. A build step walks the docs directory, concatenates the pages, splits them into overlapping chunks and embeds them into a vector store saved as a file in the project; at query time the tool retrieves the passages that match the question and returns them to the client.
- One tool that takes a plain-language question and returns the documentation passages that match it
- The whole documentation text published as a resource, for a client that would rather read it entire
- An index built and stored locally, from the docs directory you point the extractor at
- Local embeddings, so questions are answered without calling an outside service
- A container the client drives over stdio, so nothing is left running when it is idle
Docker to run the container, Python — the project pins 3.13 — to build the index, and a local copy of the gemini-cli documentation for the extract step to walk. No account, no key: the embedding model and the index both run on your machine.
One command — docker exec -i gemini-cli-mcp-container python gemini_cli_mcp.py