NOTICE: claude code is available with Anthropic's $20/mo subscription, so I consider

codemcp fully obsolete. However, there are some good design ideas (especially around

the Git-versioning scheme) that I eventually want to port into the current generation

of agentic coding clis.

codemcp

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. I like paying for my usage with a subscription plan

because it means zero marginal cost for agent actions; no more feeling

bad that you wasted five bucks on a changeset that doesn't work.

Note that if you have Claude Max ($100/mo), Claude Code can also be used

with subscription based pricing. The value proposition for codemcp is

murkier in this case (and it is definitely inferior to Claude Code in some

respects), but you can still use codemcp with Claude Max if you prefer some

of the other UI decisions it makes. (Also, it's open source, so you can

change it if you don't like it, unlike Claude Code!)

2. It's built around auto-accept by default. I want my agent to get as

far as it can without my supervision, so I can review everything in one go at

the end. There are two key things that codemcp does differently than most

coding agents: we forbid unrestricted shell, instead requiring you to

predeclare commands the agent can use in ``codemcp.toml``, and we **Git

version all LLM edits**, so you can roll back agent changes on a

fine-grained basis and don't have to worry about forgetting to commit

changes.

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. I use vim as my daily driver editor, and coding environments

that require VSCode or a specific editor are a turn off for me.

IMPORTANT: For master users - Major changes for token efficiency

To improve codemcp's token efficiency, on master I am in the process of

changing codemcp back into a multi-tool tool (instead of a single tool whose

instructions are blatted into chat when you InitProject). This means you have

to manually approve tool use. Because tool use approval is persistent across

multiple chats, I think this is a reasonable tradeoff to make, but if you

really don't like, file a bug at

refined-claude browser

extension for supporting auto-approve tool use.

Installation

I recommend this specific way of installing and using codemcp:

1. Install `uv` and install git, if they are not installed already.

2. Install claude-mcp on your browser.

This enables you to connect to SSE MCP servers directly from the website,

which means you don't need to use Claude Desktop and can easily have

multiple chat windows going in parallel. We expect this extension should

be soon obsoleted by the rollout of

Integrations. At time of

writing, however, Integrations have not yet arrived for Claude Pro subscribers.

3. Run codemcp using ``uvx --from git+https://github.com/ezyang/codemcp@prod codemcp serve``.

You can add ``--port 1234`` if you need it to listen on a non-standard port.

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: you can run codemcp remotely! If you use

Tailscale and trust all devices on your Tailnet,

you can do this securely by passing ``--host 100.101.102.103`` (replace the

IP with the Tailscale IP address of your node. This IP typically lives in

the 100.64.0.0/10 range.) WARNING: Anyone with access to this MCP can perform

arbitrary code execution on your computer, it is EXTREMELY unlikely you want to

bind to 0.0.0.0.

4. Configure claude-mcp with URL: ``http://127.0.0.1:8000/sse`` (replace the port if needed.)

5. Unfortunately, the web UI inconsistently displays the hammer icon. However, you can verify

that the MCP server is working by looking for "[MCP codemcp] SSE connection opened" in the

Console, or by asking Claude what tools it has available (it should say

tools from codemcp are available.)

If you prefer to use Claude Desktop or have unusual needs, check out INSTALL.md for

installation instructions for a variety of non-standard situations.

Usage

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"]

The ``format`` command is special; it is always run after every file edit.

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:

• Implement a new feature
• Fix failing tests
• Do a refactor

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

Philosophy

• 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.

Configuration

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."

Troubleshooting

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.

Contributing

See CONTRIBUTING.md.