Nothing is indexed until you run the build command, and nothing refreshes until you run it again — which is the trade for having the whole documentation set on disk and searchable without a hosted search dependency. Search returns five results by default, so ask for more when you are looking for a section rather than an answer.
An MCP server that fetches the `.mdx` files from the `docs/` directory of the bucketeer-io/bucketeer-docs repository, builds a keyword and full-text index over them, and answers questions from that index. Content is cached locally with SHA hashes so a rebuild only touches files that actually changed.
- Bucketeer documentation searched by relevance, returning title, URL, path, description, excerpt and score for each hit, with the result count capped by `limit` — `search_docs`
- The full markdown body of one document by the `path` returned in a search result, with its title, description and URL — `get_document`
Node.js 18+ and npm. No account and no key — the docs it indexes are public. The package is `@bucketeer/docs-local-mcp-server` (1.3.0) and runs over stdio via npx. Build the index before the first run: `npx @bucketeer/docs-local-mcp-server build-index`, and add `--force` when you want a rebuild that ignores the cache. Cached documents land in `files/docs/`, the index in `files/index/document-index.json`.
One command — npx @bucketeer/docs-local-mcp-server