Codex–Claude Council

What Codex–Claude Council MCP server does

Codex–Claude Council MCP server coordinates discussions between ChatGPT through the Codex CLI and Claude through the Claude Code CLI. Each model first develops an independent answer. They then receive one another’s responses, explain points of agreement and disagreement, revise their positions, and continue for configured rounds. One model writes the final response and the other checks it.

Both participants can work within a selected project in plan mode. They may read files, inspect project context, search the web, and run read-only commands such as git log and git diff; the README states that they cannot change project files during this process. Existing skills, plugins, MCP servers, and CLAUDE.md or AGENTS.md instructions can apply to the participants.

How it works

The MCP interface exposes tools for starting and managing council jobs:

• council_ask starts a discussion with configurable Codex and Claude models, effort levels, a synthesizer, round limit, workspace, web search, skills, and optional session IDs.
• council_debate accepts the same configuration as council_ask.
• council_join lets either Claude or Codex join an existing session, with options for the other session, model, effort, synthesizer, workspace, web search, and skills.
• council_turn handles a council turn.
• council_result retrieves a result by job_id.
• council_cancel cancels a job by job_id.

The members can start in fresh sessions, use existing Claude Code or Codex sessions, or include the session from which the request was made. Supplying a session ID allows that participant to retain the context it already knows.

Setup and configuration

The Codex–Claude Council MCP server requires Node.js 20 or newer, the Codex CLI signed in with ChatGPT, and the Claude Code CLI signed in with a Claude plan. It does not require API keys because it invokes those locally authenticated command-line tools.

The documented quick start is:

npx -y codex-claude-council install

This installation checks Node.js and both CLIs, then installs the integration for Claude Code and Codex. Restart those clients afterward so they load it. In a project, ask a question such as whether the council can explain the project and identify the first improvement to make.

Limitations and notes

Council discussions take longer than asking one model. The README’s benchmark reports roughly three to five times the time of a single model for that test set, so this approach is better suited to questions where review or disagreement is useful than to every request. The benchmark is also described as a small sample rather than proof of general accuracy.

The README documents use on Windows, macOS, and Linux, but the participating local CLIs and their own account access remain prerequisites. The documented clients are Claude Code, Claude Desktop, and Codex; the project also describes terminal use and the ChatGPT desktop app. No claim is made that agreement guarantees correctness: the council’s value is having two models challenge and check one another.