docchat-server

Version-pinned documentation retrieval as a Model Context Protocol server. Gives Claude Code / Cursor / any MCP-aware AI grounded answers from the docs of the exact library version your lockfile pins.

LLM coding assistants answer library questions from training data, biased toward the newest version. If your project pins react@18.2.0 and the latest is 19.x, asking about useEffect gets you React 19 idioms — possibly with hooks (use(), useActionState) that don't exist in 18.2. docchat-server fixes that. Index the docs at the exact versions you pin. The MCP server exposes search_docs(library, version, query) to your AI. Now the answer comes from the right docs, with hard refusal when the question is out-of-scope.

---

See it work

A Cursor Agent asks about useEffect cleanup in a React 18.2 project. It calls search_docs on docchat, gets back chunks from the React 18.2 reference, answers with version-grounded code:

Same Cursor session asked about Flask CORS (a library docchat doesn't have indexed). The server returns a clean "no docs" + the list of what IS indexed; the agent surfaces that honestly instead of hallucinating:

And here's the MCP server registered in Cursor's tool panel, showing the two tools (search_docs + list_indexed):

---

Install

uv tool install docchat-server      # puts the binary on PATH (recommended)
# or:
pip install --user docchat-server

Requirements:

• Python 3.11+
• An OPENAI_API_KEY env var (used for query + index-time embeddings)

The Qdrant vector store runs embedded — no Docker, no separate server.

---

Quickstart (3 commands + 1 config block)

export OPENAI_API_KEY=sk-...

# Index the libraries you pin. ~30–60s each, a few cents of embeddings cost.
docchat-server index react 18.2.0
docchat-server index fastapi 0.100.0
docchat-server index vue 3.4.0

# Verify
docchat-server list

Output:

Indexed collections:
  - react_18_2_0    (192 chunks)
  - fastapi_0_100_0 (38 chunks)
  - vue_3_4_0       (62 chunks)
Supported libraries: fastapi, react, vue

Then register the server with your MCP client.

Cursor

~/.cursor/mcp.json (Windows: %USERPROFILE%\.cursor\mcp.json):

{
  "mcpServers": {
    "docchat": {
      "command": "docchat-server",
      "args": ["serve"],
      "env": { "OPENAI_API_KEY": "sk-..." }
    }
  }
}

Restart Cursor. Open Settings → Tools & MCPs — docchat should appear with search_docs and list_indexed enabled.

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "docchat": {
      "command": "docchat-server",
      "args": ["serve"],
      "env": { "OPENAI_API_KEY": "sk-..." }
    }
  }
}

Fully quit Claude Desktop (tray icon → Quit, not close-window) and reopen.

Cline / Continue.dev / any other MCP-aware client

Same JSON shape under that client's MCP config. The protocol is identical.

---

The tools the AI can call

search_docs(library, version, query, api_name?, top_k?)

Returns top-K chunks from the indexed docs of the exact pinned version. Drops hits below a per-library cosine-similarity floor (tuned per-library by eval). Returns the canonical string "No relevant chunks found" when nothing clears the floor — that's a signal to the calling LLM to refuse rather than guess.

list_indexed()

Returns the (library, version) collections currently populated locally, plus the list of libraries this server's indexer knows how to populate. Useful as a session-start probe so the agent knows what's available before answering.

---

What gets indexed (built-in libraries)

Library	Versions	Source repo
React	any (defaults to main of reactjs/react.dev)	reactjs/react.dev
FastAPI	per-tag (e.g. 0.95.2, 0.100.0 fetch from the matching git tag)	tiangolo/fastapi
Vue	any (defaults to main of vuejs/docs)	vuejs/docs

FastAPI's per-tag fetching is the real demo: indexing fastapi@0.95.2 gets the Pydantic-v1-era docs; indexing fastapi@0.100.0 gets the Pydantic-v2-era docs. Two separate Qdrant collections, two different answers depending on what you pin.

Adding more libraries: extend _LIBRARY_CONFIG in library_config.py and submit a PR. Generic --repo / --paths CLI flags for arbitrary libraries land in v0.2.

---

Known limitations (v0.0.x)

• One MCP client at a time. Embedded Qdrant holds a process-level file lock on ~/.docchat-server/qdrant/. Running two MCP clients (e.g. Cursor + Claude Desktop) against the same docchat-server simultaneously will fail on the second one. Workaround: kill the other client, or switch to server-mode Qdrant (Docker) — v0.2 milestone.
• You bring your own OpenAI key. All embeddings (index-time + query-time) go through OpenAI. Local-embeddings option via sentence-transformers is v0.3 milestone.
• Indexing is CLI-only, not an MCP tool. Deliberate — exposing index_library as an MCP tool would let any connected LLM trigger arbitrary embedding cost. The user runs docchat-server index themselves.

---

Related projects

Part of a 4-project portfolio of production AI engineering:

• DocChat VS Code extension (Marketplace) — Same retrieval engine, chat panel instead of MCP tool surface.
• mneme (PyPI as smolAmem) — Multi-tier memory used by the DocChat extension's agent.
• toolpicker (PyPI) — Tool routing used by the DocChat extension's agent.

The GitHub repo for this server is still named docchat-mcp (predates the PyPI rename to docchat-server — docchat-mcp was taken on PyPI).

---

Roadmap

• v0.1 — detect_pinned_libraries(workspace_path) MCP tool, find_in_changelog(library, version, query) MCP tool
• v0.2 — --repo / --paths CLI flags for arbitrary library indexing; optional Docker-Qdrant backend for multi-client access
• v0.3 — local embeddings via sentence-transformers (drops OPENAI_API_KEY requirement for embeddings)

---

License

MIT. See LICENSE.