clone_and_write_prompt is the idea worth copying: the system prompt, the model choice and the available MCP tools all live in the repository's .agent directory, so whoever maintains that codebase decides how an agent should work on it — not whoever happens to be running the client. The clone-first design also means the agent never touches your working tree, which makes it reasonable to hand it a task and go do something else.
An MCP wrapper around OpenAI's Codex CLI. Instead of exposing file and shell primitives, it takes a repository URL and a request, clones the repository somewhere else, and runs Codex there — so the coding agent works in its own checkout rather than in yours.
- run_codex — clone a repository, optionally check out a branch and narrow to a folder, then run Codex against your request
- clone_and_write_prompt — the same, but it reads the repository's own instructions first: the system prompt from .agent/system.md and the model id from .agent/agent.json, writes the request to a .prompt file, and calls Codex with that model
- A repository can register its own MCP tools in .agent/mcps.json, which Codex loads automatically — so the agent working on that repository gets the tools that repository needs
- Runs over stdio, or over SSE on a port you choose
- A Docker setup that brings up the wrapper alongside open-responses-server, which supplies Responses API compatibility
Node 22 — that is Codex's requirement, not the wrapper's — and the Codex CLI installed globally as @openai/codex, which means an OpenAI account and its costs. The Python side installs from the repository. Everything Codex does happens in a fresh clone, so the tools need network access to your git host and credentials for any private repository.