Docs MCP Server

An MCP (Model Context Protocol) server that gives an LLM host (Claude Desktop, an IDE, etc.) a tool for pulling live documentation, instead of relying on the model's training data. Searches are scoped to a library's official docs site, fetched pages are chunked and embedded into a local vector store, and the tool returns only the most relevant excerpts — not raw scraped HTML — for the query.

How it works

LLM host (e.g. Claude Desktop)
        │  calls get_docs(query, library) over MCP (stdio)
        ▼
main.py
   │  1. site-scoped Google search via Serper API
   │  2. fetches matching pages concurrently (httpx + asyncio.gather)
   │  3. strips HTML → plain text (BeautifulSoup)
   ▼
rag.py
   │  4. chunks each page (langchain-text-splitters)
   │  5. embeds + upserts chunks into a local Chroma collection,
   │     scoped by library, keyed by URL (idempotent — re-fetching
   │     a page doesn't duplicate it)
   │  6. queries the collection for the top-k chunks most relevant
   │     to the original query
   ▼
Returns the top-k excerpts (with source URLs) back to the LLM host

Embeddings run locally via Chroma's bundled model — no external embedding API key required.

Project structure

File	Responsibility
main.py	MCP server entry point; defines the get_docs tool; search + fetch logic
rag.py	Chunking, embedding, and retrieval against a local Chroma vector store
web.py	Minimal FastAPI dashboard for exercising get_docs from a browser (no MCP host needed)
tests/	pytest suite; all HTTP calls mocked via respx, vector store tests use an in-memory Chroma client
.github/workflows/ci.yml	Lint (ruff) + test on every push/PR

Requirements

• Python 3.11+
• uv package manager
• A Serper API key (free tier available)

Setup

uv sync --all-groups
cp .env.example .env   # then add your SERPER_API_KEY

Running the MCP server

uv run main.py

This starts the server over stdio — it's meant to be launched by an MCP host, not run standalone for interactive use (see below).

Running the web dashboard

To try the tool from a browser instead of an MCP host:

uv run uvicorn web:app --reload

Then open http://127.0.0.1:8000.

Connecting to Claude Desktop

Edit your Claude Desktop config (claude_desktop_config.json):

{
    "mcpServers": {
        "docs": {
            "command": "uv",
            "args": [
                "--directory",
                "/ABSOLUTE/PATH/TO/THIS/PROJECT",
                "run",
                "main.py"
            ]
        }
    }
}

Restart Claude Desktop, then ask something like "how do I use Chroma DB with LangChain?" — Claude will call get_docs and answer from the live docs instead of its training data.

Supported libraries

Currently langchain, llama-index, and openai — see DOCS_URLS in main.py. Add your own by adding an entry mapping a library name to its docs site.

Testing

uv run pytest -v
uv run ruff check .

License

MIT — see LICENSE.