llmdocs-mcp 0.2.1

Agent-first documentation platform (CLI and server)

llmdocs

Documentation your agents can actually search.

Self-hosted stack: hybrid search (semantic + keyword), MCP tools, raw .md URLs, and llms.txt — no SaaS, no paid vector API required.

📦 PyPI → llmdocs-mcp  ·  🐳 Docker Hub

pip install llmdocs-mcp · Python 3.12+ · Embedded Chroma · Local or OpenAI-compatible embeddings

Agents: add hosted docs MCP → llmdocs-production.up.railway.app/mcp/ · setup

Get Started · Why llmdocs? · Features · CLI · MCP · Agents (hosted MCP) · HTTP · Docker · Docs

---

Your docs are buried in a repo. Your agent can’t find them.

Paste links into chat and hope for the best? Copy whole folders into context? That doesn’t scale — and RAG pipelines you don’t own are a second job.

llmdocs indexes your Markdown (with frontmatter), serves hybrid search over real sections, and exposes MCP so Cursor, Claude, and custom clients can search_docs, get_doc, and list_docs against your corpus. You run the server; embeddings stay local by default.

---

Why llmdocs?

llmdocs	Typical “docs in chat”
Indexed sections (H2/H3 chunks) + BM25	Giant paste or ad-hoc grep
Semantic + keyword fusion	Embeddings only, or keywords only
Stable URLs — GET /guide.md raw body	Screenshots or broken links
MCP at /mcp/ (Streamable HTTP)	No standard tool surface
Self-hosted, local or API embeddings	Another vendor’s vector bill

---

Get started

A few commands. Install from PyPI (package name llmdocs-mcp; the program is llmdocs).

Step	Action
1	pip install llmdocs-mcp
2	llmdocs init — detects an existing docs/ (or doc/, documentation/) and generates llmdocs.yaml. Creates a sample docs/index.md only if no docs folder is found. Use --docs-dir PATH to point at a specific folder.
3	llmdocs build — index for search + MCP
4	llmdocs serve — HTTP + MCP (use llmdocs watch in another terminal for auto-rebuild on save)

pip install llmdocs-mcp
llmdocs init
llmdocs build
llmdocs serve

From a git checkout (minimal run):

python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt && pip install .
llmdocs --help

From a git checkout (development):

python -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt && pip install -e .
pytest

Note: The CLI is the supported user interface. llmdocs is not a stable Python SDK — internal modules are implementation details.

---

What you get

• Hybrid search — Chroma semantic vectors + BM25 keyword fusion over chunked sections.
• MCP (FastMCP, Streamable HTTP) — search_docs, get_doc, list_docs at POST /mcp/ (use a trailing slash in client URLs).
• Raw markdown — GET /<path>.md returns body without YAML frontmatter.
• llms.txt — Generated on llmdocs build and served at GET /llms.txt.
• CLI — init, serve, build, validate, search, watch, and more.

Default embedding model: sentence-transformers/all-MiniLM-L6-v2 (local, no API key needed). Switch to OpenAI, Azure OpenAI, or any OpenAI-compatible proxy (LiteLLM, vLLM) by setting embeddings.provider: openai in llmdocs.yaml — see Configuration.

---

CLI

llmdocs --version
llmdocs init [--config llmdocs.yaml] [--docs-dir PATH] [--force]
llmdocs build --config llmdocs.yaml
llmdocs serve --config llmdocs.yaml [--host HOST] [--port PORT]
llmdocs validate --config llmdocs.yaml
llmdocs search "your query" --config llmdocs.yaml
llmdocs watch --config llmdocs.yaml   # rebuild on file changes

Full CLI reference →

---

MCP tools

Tool	What it does
search_docs	Hybrid semantic + keyword search over indexed chunks
get_doc	Full document body + metadata by path
list_docs	List documents (optional category / path prefix filters)

MCP reference → · Quickstart (local URL) →

Hosted llmdocs MCP (for agents)

Give Cursor, Claude Code, VS Code, and other MCP clients searchable access to this project’s official docs (install, CLI, Docker, Railway, etc.) — no local llmdocs serve required.

MCP URL	https://llmdocs-production.up.railway.app/mcp/ (keep the trailing slash)
Health	/health

Cursor — add to ~/.cursor/mcp.json or .cursor/mcp.json:

{
  "mcpServers": {
    "llmdocs-docs": {
      "name": "llmdocs",
      "url": "https://llmdocs-production.up.railway.app/mcp/"
    }
  }
}

Claude Code: claude mcp add --transport http llmdocs-docs https://llmdocs-production.up.railway.app/mcp/ (optional: --scope project for a shared .mcp.json).

VS Code (.vscode/mcp.json): "servers": { "llmdocs-docs": { "type": "http", "url": "https://llmdocs-production.up.railway.app/mcp/" } }.

This is a public, read-only docs index (same content as the open docs/ folder). For your private documentation, deploy llmdocs yourself and point clients at https://<your-host>/mcp/.

Full instructions and self-hosted URLs: docs/guides/mcp-clients.md.

MCP clients (generic)

For local dev, point clients at http://localhost:8080/mcp/ (or your llmdocs.yaml host/port). llmdocs uses Streamable HTTP everywhere.

Product	Docs
Claude Code	Anthropic: Claude Code + MCP
Cursor	Cursor MCP
VS Code	VS Code: MCP servers

---

HTTP surface

Route	Description
GET /	JSON metadata and endpoint index
GET /health	Liveness JSON ({"status": "healthy"})
POST /mcp/	Streamable HTTP MCP (trailing slash)
GET /<path>.md	Raw markdown body (no frontmatter)
GET /llms.txt	Generated catalog (after llmdocs build)

---

Docker

Image: vinny2prg/llmdocs-mcp

docker run --rm -p 8080:8080 \
  -v "$(pwd)/docs:/docs:ro" \
  -v llmdocs_data:/data \
  vinny2prg/llmdocs-mcp:latest

The repo includes docker-compose.yml and a Railway guide — mount a volume on /data for persistent Chroma + generated files.

---

Development

python -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt
pip install -e .
pytest

Git: This repository is Projects/llmdocs. If your editor opened a parent folder that is also a Git repo, run git commands from the project root so commits apply only to llmdocs.

---

Project layout (high level)

llmdocs/
  cli.py           Click CLI
  config.py        llmdocs.yaml (Pydantic v2)
  server.py        FastAPI + lifespan indexing + /mcp mount
  indexing/        Parser, chunker, Chroma indexer, hybrid search
  mcp/             FastMCP tools + runtime

---

Documentation

Doc	Description
Installation	PyPI, Docker, from source
Quickstart	Config, index, browser & MCP
MCP clients	Cursor, VS Code, Claude Code
Configuration	llmdocs.yaml reference
Hosting	Docker, TLS, reverse proxy
Railway	GitHub deploy, volumes, PORT
CLI reference	All commands and flags
Contributing	Issues, PRs, style

Preview the in-repo docs site locally:

llmdocs serve --config docs/llmdocs.yaml --data-dir .llmdocs/project-docs

Run llmdocs build --config docs/llmdocs.yaml first so search works.

---

License

MIT