code-quality-mcp

Stop asking your LLM if your Python is clean. Let it actually check.

An MCP server that gives Claude, Cursor, VS Code (or any MCP-compatible client) a real, deterministic view of your Python code quality — powered by the tools you already trust: flake8, mypy, McCabe, and vulture.

No more "looks good to me" from an LLM that never ran a linter. This is ground truth.

Why

LLMs hallucinate code reviews. They miss style violations, invent complexity numbers, and confidently declare your dead code "actively used."

code-quality-mcp plugs the gap: it runs the canonical Python quality tools and hands the LLM a clean, structured report. The model stops guessing and starts citing.

It also ships with a hard instruction baked into the server: never modify, refactor, or fix your code unless you explicitly ask. It reports. You decide.

Related MCP server: MCP Code Checker

What it gives you

Reports come back as markdown tables — readable for you, parseable for the model.

Install

git clone https://github.com/Javier-Morenosa/code-quality-mcp.git
cd code-quality-mcp
pip install -e .

Requires Python 3.10+.

Wire it up

Cursor

Add to your Cursor MCP settings (or copy cursor_mcp.json):

{
  "mcpServers": {
    "code-quality": {
      "command": "python",
      "args": ["-m", "code_quality_mcp.server"],
      "env": {
        "PYTHONPATH": "C:/path/to/code-quality-mcp/src"
      }
    }
  }
}

VS Code

See vscode_mcp.json for a ready-to-use config.

Claude Code

claude mcp add code-quality -- python -m code_quality_mcp.server

Use it

Once the server is wired up, just talk to your LLM:

"Run a full quality analysis on src/my_module.py."

"Is there dead code in this project?"

"Which functions are too complex?"

The model will call the right tools, get real output from real linters, and give you a report grounded in facts — not vibes.

Pro tip

Call set_workspace once at the start of a session to point the server at your project root. After that, you can pass relative paths and everything Just Works.

What it will NOT do

By design, this server is read-only. It will not:

• Edit your files

• Refactor your code

• Auto-fix style issues

• Suggest changes unless you explicitly ask for them

If you want fixes, run black, ruff --fix, or ask your LLM directly. This server's job is to tell the truth about your code, nothing more.

Roadmap

• ruff backend as a faster drop-in for flake8

• Per-file complexity heatmap

• Configurable rule sets via pyproject.toml

License

MIT. Use it, ship it, fork it.