MCP

codex-cc-orchestrator: Dispatch Tasks from Codex to Claude Code for Execution and Review

The dispatch station is best suited for having Codex dispatch coding tasks to Claude Code for execution and review by an independent read-only session — but only if you explicitly pass in a trusted project_root and have both Agents installed.

Official source · 点将台 codex-cc-orchestrator

Install

git clone https://github.com/zjgxkj/codex-cc-orchestrator.git && cd codex-cc-orchestrator && git checkout v0.4.0 && uv sync

The dispatch station (codex-cc-orchestrator) is for scenarios where a controlling Agent needs to dispatch tasks to an executor and then have them reviewed by an independent read-only session. The most important point before you start: you must have both local Agents, Codex and Claude Code, installed and configured, and every task must explicitly pass in a trusted project_root — the service will not guess your workspace from the current directory.

Who it's for

  • Users already using Codex who have installed Claude Code and want to dispatch tasks to Claude Code for execution and review via MCP, enabling layered collaborative orchestration.
  • Development workflows that need task dispatch, execution, resumption, independent review, result compression, and usage tracking.
  • Not suitable for: anyone without both Codex and Claude Code environments set up. The evidence does not state whether either side can be used alone — defer to the official documentation.
  • Prerequisites: Python 3.13+ and uv on your machine. Full Chinese installation steps are in the official INSTALL.md.

First attempt in 3 minutes

Prerequisites: Python 3.13+, uv, and configured Codex and Claude Code.

  1. Get and install (official main path):
git clone https://github.com/zjgxkj/codex-cc-orchestrator.git
cd codex-cc-orchestrator
git checkout v0.4.0
uv sync
  1. Register the service in Codex's STDIO MCP configuration (official Windows example):
[mcp_servers.codex-claude-agent]
command = 'D:\path\to\codex-claude-agent-mcp\.venv\Scripts\python.exe'
args = ['-m', 'codex_claude_agent_mcp']
enabled = true
startup_timeout_sec = 30
tool_timeout_sec = 7200
  1. Do a low-risk first test: the service exposes a ping tool; use it to confirm the MCP channel works, with no file writes involved. A success signal is that Codex can list the service's tools and receive the ping response. Note that stdout carries only MCP protocol messages — logs and errors go to stderr, so check stderr when troubleshooting startup issues.

Actually running tasks (execute_task etc.) is a different matter: you must explicitly pass in a project_root you trust, and execution will read and write files within that directory. This is not a read-only operation — confirm the target project directory before calling.

Use cases

Case 1: Dispatch and execute a coding task. In a Codex session, call execute_task with the task description and the project_root you explicitly authorize. The executor session uses a restricted toolset (Read/Grep/Glob/Edit/Write/Bash — network access allowed, nested Agents and notebook editing forbidden) to make the changes. What to expect: the task is executed and its state persisted to local SQLite (~/.codex-claude-agent-mcp/state.db).

Case 2: Independent review and resumption. Calling review_task starts a brand-new read-only session that reviews the execution result and returns PASS/FAIL; if FAIL, you can call continue_task to continue from the saved executor session with feedback included. What to expect: the read-only review does not modify files, and resumption reuses the session context. You can also use run_job / run_jobs to orchestrate multiple tasks and get_job_status to check progress.

Boundaries of use

  • Read/write nature: ping is a read-only liveness probe; execute_task, run_job, and run_jobs write files under project_root and have real side effects — verify the directory is correct before calling; review_task performs review in a read-only session.
  • Authorization: each task can receive a project_root, and Codex must explicitly pass in a trusted workspace root; the service will not guess from its own process cwd. ALLOWED_PROJECT_ROOTS is an optional long-term trust list.
  • Local state: executor sessions and other state are persisted in ~/.codex-claude-agent-mcp/state.db. The official documentation does not describe a deletion or revocation path for this state.
  • Concurrency limit: CLAUDE_MAX_CONCURRENCY (default 4) applies only to a single MCP server process.
  • A tool's read-only hint only means it performs no write operations — it does not imply no network access or an absolute privacy boundary. Unspecified aspects in the official docs (such as credential management details) should not be assumed risk-free.