Scoping searches to four known documentation sites is why answers come back on-target instead of drowning in blog posts and Stack Overflow copies. The flip side is equally plain: the library list lives in the source, so covering a fifth framework is a code change, not a config change. Worth it if those four are the ones you live in.
An MCP server that answers documentation questions against the live docs rather than a model's memory. It takes a natural-language query, runs a site-specific search through the Serper API, then parses the resulting pages with BeautifulSoup and returns the documentation text.
- Documentation search in natural language, scoped to one library rather than the open web
- Coverage of Langchain, Llama-Index, MCP and OpenAI out of the box, with more added by editing the source
- Page content parsed and returned as readable text, not a list of links
- Current documentation, which is the point for models whose training predates the last few releases
A Serper API key, since every search goes through Serper. The server runs through uv with the MCP CLI extra — the documented client config invokes `uv run --with mcp[cli] mcp run server.py` against your local checkout, so this is a clone-and-run rather than a package install. The project is `docs-mcp-server` (0.1.0 in pyproject). Adding a library beyond the four supported ones means editing the code.