Clean MCP

Local semantic code search for AI coding agents — runs on your laptop, indexes stay on disk.

Clean MCP is an MCP server that gives Claude Code, Cursor, and other AI tools meaning-aware code search. It parses your repositories with tree-sitter, builds a call graph, embeds every function with a local sentence-transformer model, and stores everything in LanceDB — no cloud, no API keys, no telemetry.

"find the function that validates email on signup"
   ↓
search_code(query="email validation on signup")
   ↓
returns the right function with full source + callers/callees

Features

• Semantic search — describes behaviour, not keywords; finds code by what it does.
• Local-only — embeddings, metadata, and source files live in ~/.clean/. Nothing leaves your machine.
• MCP-native — drops into Claude Code / Cursor / any MCP client over stdio.
• Index anything — point it at a local folder or a public GitHub repo.
• Tree-sitter parsing — Python, JavaScript, TypeScript.
• Call graph aware — search results include direct callers and callees.

Installation

Requires Python 3.10–3.13.

git clone https://github.com/cleanmcp/clean-mcp.git
python -m pip install -e ".[dev]"

This project uses a src/ layout, so the module command only works after the package is installed into the Python environment you are using. If you create or activate a virtualenv, run python -m pip install -e ".[dev]" again inside that environment before starting the server. The dev extra installs the project’s local tooling, including pytest and ruff.

First search or index downloads all-MiniLM-L6-v2 (~90 MB) into ~/.cache/huggingface/ if it is not already cached. After that, the model is reused from disk and does not download again.

The server starts with heavy dependencies lazy-loaded. Launching clean should reach Starting Clean MCP server quickly; the embedding model and LanceDB stack load only when you call search_code or index_repo for the first time in that process.

Running the server

Clean is a stdio MCP server. It speaks the Model Context Protocol over stdin/stdout — there is no HTTP port, no web framework, and nothing to run under uvicorn/gunicorn. In normal use your MCP client (Claude Code, Cursor, …) launches the process for you based on the config below; you don’t start it by hand.

To run it manually — for debugging, or to confirm it boots — use either of these equivalent commands:

# Module form
python -m clean.local.mcp_server

# Console script (installed by `python -m pip install -e ".[dev]"`)
clean

The process then waits silently for an MCP client to connect over stdin/stdout. That silence is expected — it is not a web server and will not print a URL. Do not type into that terminal; blank lines or other text are treated as MCP input and will produce JSON-RPC parse errors. Press Ctrl+C to stop it. To talk to it interactively, use the MCP Inspector:

npx @modelcontextprotocol/inspector python -m clean.local.mcp_server

Pin a default repo

Pass --repo owner/repo so search_code targets that repo without callers specifying it each time:

python -m clean.local.mcp_server --repo facebook/react

Where it stores data

Override the on-disk locations with environment variables (see Where your data lives):

CLEAN_PERSIST_PATH=/data/clean/index python -m clean.local.mcp_server

Wire it into your MCP client

Claude Code / Cursor / Codex — .mcp.json (project-scoped)

{
  "mcpServers": {
    "clean": {
      "command": "python",
      "args": ["-m", "clean.local.mcp_server"]
    }
  }
}

Or globally via the CLI:

# Claude Code
claude mcp add clean -- python -m clean.local.mcp_server

# Codex
codex mcp add clean -- python -m clean.local.mcp_server

Tools

Indexing modes

Ask your MCP client to index code in one of two ways:

• Index this local directory

  • Calls index_repo with path.
  • Indexes a folder already on disk.
  • Does not clone anything.
  • Uses the folder basename as the repo name unless it can detect a GitHub remote, in which case it uses owner/repo.
  • Starts in the background by default; use list_repos to check for ready.

• Index this GitHub repo owner/repo

  • Calls index_repo with repo.
  • Expects a GitHub repo in owner/repo format.
  • Uses RepoManager to clone or locate the repo under the configured repos directory, usually ~/.clean/repos.
  • Then indexes that checked-out local copy.
  • Starts clone and indexing in the background by default; use list_repos to check for ready.

Useful prompt variants:

Index this local directory
Index this GitHub repo clarsbyte/obs-assistant
Force re-index this GitHub repo clarsbyte/obs-assistant
Index the main branch of this repo in the foreground

Example usage in Claude Code

“Index this directory” → calls index_repo with the current path

“Find the function that handles login redirects” → search_code

“Show me how the indexer entry point works” → search_code then get_source

Where your data lives

~/.clean/
├── index/         LanceDB vector store
├── metadata.db    SQLite — which repos are indexed, status
└── repos/         git clones (only for GitHub-mode indexing)

Back up that folder to keep your indexes. Delete it to start fresh.

Override the location with env vars:

Development

make install   # creates .venv and installs deps
make test      # runs the test suite
make lint      # ruff check + format check
make format    # apply ruff fixes

How the token savings work

Clean is engineered to minimise the tokens an AI agent spends understanding your code — via semantic retrieval, tiered result summaries, on-demand expansion, and incremental indexing. For a full, code-referenced breakdown of every cost-reduction mechanism, see docs/cost-reduction.md.

Contributing

PRs welcome. See CONTRIBUTING.md.

License

MIT — see LICENSE.