Docs Search MCP Server

Give Claude, Cursor or any MCP client a search tool over your own documentation folders: internal docs, a repo's docs/ folder, a vendor's docs pulled from its llms.txt. The agent finds the right section, reads it and cites path#section, instead of guessing from training data or pasting whole files into the chat.

• Formats: Markdown, MDX, reStructuredText, HTML and llms.txt. Files are split at each heading, and every section keeps its heading trail (Webhooks > Retries).
• Search: BM25 ranking with heading and phrase boosts. It understands identifiers like invoice.payment_failed. No embeddings, no API keys, no database.
• Always fresh: when files are added or edited, the index rebuilds on the next search.
• Read-only: 5 tools, all marked read-only. The server can only read files it indexed under the folders you give it.
• Pull a docs site: python3 -m docsmcp pull https://docs.example.com --out ~/docs/example downloads its llms.txt and the Markdown pages it links to, same host only.
• Python 3.9+, standard library only. Speaks MCP 2026-07-28 (stateless) and the 2025-11-25 handshake.

Claude Code

claude mcp add docs-search -- python3 /path/to/docs-search-mcp-server/server.py --docs /path/to/docs --docs /path/to/more/docs

Or install it as a plugin (/plugin marketplace add, then point DOCS_MCP_DIRS at your folders, separated by :, or ; on Windows). That also adds the docs-search skill, which tells Claude to search, read and cite.

Claude Desktop, Cursor and others: copy examples/configs/claude_desktop_config.json or cursor-mcp.json and set the two paths.

$ python3 -m docsmcp search examples/docs "how do I verify webhook signature"
    5.39  guides/webhooks.md#verify-signatures  (Webhooks > Verify signatures)
$ python3 -m docsmcp read examples/docs reference/rate-limits.rst handling-429-errors

./demo.sh starts the server and runs an MCP client against the example docs (a fictional billing API). ./run_tests.sh runs 14 tests, including protocol tests and an llms.txt pull from a local test server.

• Skipped: .git, node_modules, virtualenvs, build folders, hidden folders, and files over 5 MB.
• It suits up to tens of thousands of sections. Very large doc sets work but take a few seconds to index at start.
• pull fetches one page at a time, with a short pause, at most --max pages (default 200). Links in the llms.txt "Optional" section are fetched only with --optional. Follow each site's terms of use.
• The text of the docs is passed to the model as data with a note not to follow instructions in it. Only index folders you trust.

Related: Python MCP Server Template (free) for building your own server: https://youraitools.co/marketplace/python-mcp-server-template. GitHub Issue Triage MCP Server: https://youraitools.co/marketplace/github-issue-triage-mcp

• Search is BM25 keyword ranking with heading and phrase boosts.
• No embeddings, no API keys, no database.
• The server can only read files it indexed under the folders you give it.
• Only index folders you trust.