Codex Plugin for Claude Code
OpenAI's official Codex plugin for Claude Code provides eight slash commands for read-only code review, adversarial review, task delegation, rescue work, background job management, and session transfer while reusing the local Codex CLI's authentication and configuration.
Project overview
Codex Plugin for Claude Code is OpenAI's official Claude Code plugin for adding independent code review, adversarial review, rescue tasks, background jobs, and session transfer through the locally configured Codex CLI. It is not a hosted agent or standalone IDE: Claude Code provides the entry point and context, while the plugin starts local Node.js and Codex app-server processes. Authentication, models, endpoints, and repository access follow the user's Codex configuration. Normal review is designed for read-only analysis, while rescue can modify a project within its authorized scope, so those permissions need to remain distinct.
Core capabilities
Read-only code review
/codex:review asks Codex to inspect the current working tree or a specified scope. Its command constrains the task to read-only behavior and permits Read, Glob, Grep, and limited Node or Git calls.
Steerable adversarial review
/codex:adversarial-review maintains a session that can be challenged and followed up, adding an independent perspective for risky designs or fixes.
Rescue and task delegation
/codex:rescue invokes the codex-rescue agent to investigate or complete work. Unless the user explicitly asks for read-only diagnosis, the agent may write files and run commands.
Background job control
Status, result, and cancel inspect, retrieve, and stop background Codex work. Job state and logs are stored locally and partitioned by workspace.
Claude session transfer
/codex:transfer accepts only path-validated JSONL records under ~/.claude/projects and creates visible, resumable turns in the Codex app or TUI.
Installation and usage
Add the OpenAI marketplace in Claude Code, install codex@openai-codex, reload plugins, and run /codex:setup. Setup checks or helps install the local Codex CLI. Global npm installation, authentication, and enabling the Stop review gate should each receive separate approval.
Let an AI Agent install it
Send this prompt to Codex, Claude Code, or another AI agent that can work with your local environment.
Help me install Codex Plugin for Claude Code from https://github.com/openai/codex-plugin-cc. First review the README, v1.0.6 release, LICENSE, plugin manifest, hooks, and setup command, then check the current Claude Code version, Node.js version, and local Codex CLI state. Explain that it starts local processes, reads the repository and Codex configuration, stores job logs, may write through rescue, and can read ~/.claude/projects sessions through transfer. After I approve, run the README's marketplace, install, and reload commands. Ask again before any global npm installation, Codex login, API key entry, extra permission, config overwrite, or review-gate enablement, and never print credentials. Finally, run only a background read-only review in a non-production repository and verify it with status and result. Do not run rescue or transfer.- Claude Code with plugin marketplace support
- Node.js 18.18 or newer
- A ChatGPT subscription or OpenAI API key for local Codex CLI authentication
- A safe test repository and a decision about whether rescue may write
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup- 1Add the OpenAI marketplace
Run /plugin marketplace add openai/codex-plugin-cc in Claude Code, then /plugin install codex@openai-codex.
- 2Reload and complete local setup
Run /reload-plugins and /codex:setup. Check Node.js, the Codex CLI, and authentication before approving global installation or login changes.
- 3Run a read-only background review
In a non-production repository, run /codex:review --background with a narrow scope and without secrets, personal data, or unrelated files.
- 4Retrieve and verify the result
Use /codex:status to inspect the job, then /codex:result to retrieve it. Treat findings as review advice that still needs human verification.
- 5Open higher-risk capabilities selectively
Use rescue or transfer only after reviewing write and data boundaries. Enable the review gate only while someone actively monitors it.
After reload and setup, confirm that /codex:* commands and the codex-rescue agent are visible. In a non-production repository, run /codex:review --background, followed by /codex:status and /codex:result. Do not use rescue, transfer, or the review gate for the first verification.
- Never paste an API key into chat, draft content, or repository files
- Setup may suggest npm install -g @openai/codex, which is a global environment change requiring approval
- The review gate is optional and can create long loops, so leave it disabled during initial setup
Use cases
Independent pre-commit review
Have Codex inspect the current diff from a perspective separate from Claude Code, with a developer verifying every finding.
Challenge a risky design
Use adversarial review to probe edge cases, assumptions, and missing tests without automatically changing the implementation.
Delegate a bounded repair
Send a narrow implementation or debugging task to rescue while constraining its impact with a branch, sandbox, and command permissions.
Continue work across clients
Import a selected Claude Code session into the Codex app or TUI as visible turns that can be continued.
Assessment
The plugin's value is a traceable two-agent workflow between Claude Code and local Codex, not a promise that either model guarantees quality. The official OpenAI repository provides explicit installation commands, an Apache-2.0 license, versioned releases, command source, hooks, and tests. The repository also makes the normal review's read-only contract distinguishable from rescue's write-capable default. The cost is a wider data and permission surface: repository content and prompts are processed through the endpoint configured for local Codex, transfer reads Claude sessions, background jobs write local logs, and the review gate can extend loops and consume usage. A measured rollout starts with setup and read-only review in a non-production repository, then selectively enables rescue, transfer, or the Stop hook. This assessment is based on first-party public material and source structure; installation and behavior were not tested.
Why it may be useful
- Official repository, explicit installation path, versioned releases, and Apache-2.0 license
- Eight commands covering review, delegation, job management, and session transfer
- Clear distinction between read-only review and write-capable rescue instructions
- Automated tests, hooks, and actual runtime scripts rather than a placeholder repository
What to know first
- Depends on Claude Code, Node.js, Codex CLI authentication, and available usage
- Rescue, transfer, and the review gate expand file, transcript, and model data boundaries
- The CHANGELOG trails the releases and the repository has no dedicated SECURITY.md
- No plugin command was installed or executed for this review
README
Codex Plugin for Claude Code
Overview
OpenAI's official Codex plugin for Claude Code provides eight slash commands for read-only code review, adversarial review, task delegation, rescue work, background job management, and session transfer while reusing the local Codex CLI's authentication and configuration. Use Codex from Claude Code to review code or delegate tasks.
Getting started
- Add the OpenAI marketplace in Claude Code, install codex@openai-codex, reload plugins, and run /codex:setup. Setup checks or helps install the local Codex CLI. Global npm installation, authentication, and enabling the Stop review gate should each receive separate approval.
- The README describes a local bridge between Claude Code and Codex for normal or adversarial review and for delegated Codex subagent work.
- Installation adds the openai/codex-plugin-cc marketplace, installs codex@openai-codex, reloads plugins, and runs /codex:setup.
- After reload and setup, confirm that /codex:* commands and the codex-rescue agent are visible. In a non-production repository, run /codex:review --background, followed by /codex:status and /codex:result. Do not use rescue, transfer, or the review gate for the first verification.
Configuration
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setupRead the complete README on GitHub →