Make Claude Desktop a pair programming assistant by installing codemcp. With it, you can directly ask Claude to implement features, fix bugs and do refactors on a codebase on your computer; Claude will directly edit files and run tests. Say goodbye to copying code in and out of Claude's chat window!

codemcp offers similar functionality to other AI coding software (Claude Code, Cursor, Cline, Aider), but it occupies a unique point in the design space:

1. It's intended to be used with Claude Pro, Anthropic's $20/mo subscription offering. Say goodbye to giant API bills. (Say hello to time-based rate limits.)

2. It's built around safe agentic AI by providing a limited set of tools that helpful, honest and harmless LLMs are unlikely to misuse, and enforcing best practices like use of Git version control to ensure all code changes can be rolled back. As a result, you can safely unleash the AI and only evaluate at the end if you want to accept the changes or not.

3. It's IDE agnostic: you ask Claude to make changes, it makes them, and then you can use your favorite IDE setup to review the changes and make further edits.

First, install uv and install git, if they are not installed already (on Windows, if you installed Git, I recommend rebooting).

Then, in claude_desktop_config.json:

{
  "mcpServers": {
    "codemcp": {
      "command": "/Users/<username>/.local/bin/uvx",
      "args": [
        "--from",
        "git+https://github.com/ezyang/codemcp@prod",
        "codemcp"
      ]
    }
  }
}

On Windows, double backslashes are necessary for the path:

C:\\Users\\<username>\\.local\\bin\\uvx.exe

If the MCP successfully loaded, a hammer icon will appear and when you click it "codemcp" will be visible.

Pro tip: If the server fails to load, go to Settings > Developer > codemcp > Logs to look at the MCP logs, they're very helpful for debugging.

Pro tip: if on Windows, the logs say "Git executable not found. Ensure that Git is installed and available", and you just installed Git, reboot your machine (the PATH update hasn't propagated.) If this still doesn't work, open System Properties > Environment Variables > System variables > Path and ensure there is an entry for Git.

Pro tip: if you like to live dangerously, you can change prod to main. If you want to pin to a specific release, replace it with 0.3.0 or similar.

Pro tip: it is supported to specify only uvx as the command, but uvx must be in your global PATH (not just added via a shell profile); on OS X, this is typically not the case if you used the self installer (unless you installed into a system location like /usr/local/bin).

First, you must create a codemcp.toml file in the Git repository checkout you want to work on. If you want the agent to be able to do things like run your formatter or run tests, add the commands to execute them in the commands section (note: these commands need to appropriately setup any virtual environment they need):

format = ["./run_format.sh"]
test = ["./run_test.sh"]

Next, in Claude Desktop, we recommend creating a Project and putting this in the Project Instructions:

Initialize codemcp with $PROJECT_DIR

Where $PROJECT_DIR is the path to the project you want to work on.

Then chat with Claude about what changes you want to make to the project. Every time codemcp makes a change to your code, it will generate a commit.

To see some sample transcripts using this tool, check out:

codemcp will generate a commit per chat and amend it as it is working on your feature.

• When you get rate limited, take the time to do something else (review Claude's code, review someone else's code, make plans, do some meetings)

• This is not an autonomous agent. At minimum, you have to intervene after every chat to review the changes and request the next change. While you can ask for a long list of things to be done in a single chat, you will likely hit Claude Desktop's output limit and have to manually "continue" the agent anyway. Embrace it, and use the interruptions to make sure Claude is doing the right thing.

• When Claude goes off the rails, it costs you time rather than dollars. Behave accordingly: if time is the bottleneck, watch Claude's incremental output carefully.

Here are all the config options supported by codemcp.toml:

project_prompt = """
Before beginning work on this feature, write a short haiku.  Do this only once.
"""

[commands]
format = ["./run_format.sh"]
test = ["./run_test.sh"]

The project_prompt will be loaded when you initialize the project in chats.

The commands section allows you to configure commands for specific tools. The names are told to the LLM, who will decide when it wants to run them. You can add instructions how to use tools in the project_prompt; we also support a more verbose syntax where you can give specific instructions on a tool-by-tool basis:

[commands.test]
command = ["./run_test.sh"]
doc = "Accepts a pytest-style test selector as an argument to run a specific test."

To run the server with inspector, use:

PYTHONPATH=. mcp dev codemcp/__main__.py

Logs are written to ~/.codemcp/codemcp.log. The log level can be set in a global configuration file at ~/.codemcprc:

[logger]
verbosity = "INFO"  # Can be DEBUG, INFO, WARNING, ERROR, or CRITICAL

Logging is not configurable on a per project basis, but this shouldn't matter much because it's difficult to use Claude Desktop in parallel on multiple projects anyway.

See CONTRIBUTING.md.