Claude Code and Codex Collaboration Guide

The Codex plugin for Claude Code lets you review code and delegate tasks to Codex without leaving Claude Code. A practical division of work is:

Claude Code: requirements, planning, and design discussion

Codex: implementation or focused investigation through /codex:rescue

Codex: independent review through /codex:review

Actual cost and usage depend on your Anthropic, OpenAI, and AnyRoute plans. Monitor both sides instead of assuming one model is always cheaper.

This guide assumes Claude Code and Codex CLI are already installed and configured.

Install the Codex plugin

Run these commands separately inside Claude Code:

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup

/codex:setup checks whether Codex is installed and authenticated. If needed:

!codex login

The plugin uses the Codex CLI installation, authentication state, and configuration already present on the machine.

Official project: openai/codex-plugin-cc

Core commands

Read-only reviews

/codex:review
/codex:review --base main
/codex:review --background
/codex:adversarial-review --background look for race conditions and challenge the caching strategy

/codex:review is read-only. Use adversarial review to pressure-test a particular decision or risk.

Delegate work

/codex:rescue investigate why the tests started failing
/codex:rescue fix the failing test with the smallest safe patch
/codex:rescue implement validation for the registration form
/codex:rescue --background refactor the database connection pool
/codex:rescue --model gpt-5.4-mini write unit tests for utils.ts
/codex:rescue --model spark fix this lint error
/codex:rescue --resume apply the preferred fix from the previous run

If no model or reasoning effort is specified, Codex chooses its configured defaults. The spark alias maps to gpt-5.3-codex-spark in the current plugin.

Manage background jobs

/codex:status
/codex:result
/codex:cancel

When available, /codex:result includes a session ID that can be continued directly:

Transfer the current Claude Code session

/codex:transfer

This creates a persistent Codex thread from the current Claude Code session when the installed versions support the external-agent importer.

Recommended workflow

Normal development

Requirement → Claude plans → Codex implements → Codex reviews → developer verifies

Describe the requirement and constraints.

Ask Claude to identify the implementation plan and risks.

Delegate clearly bounded implementation tasks with /codex:rescue.

Run /codex:review --background.

Evaluate the findings, fix confirmed issues, and run project tests.

Codex does not automatically inherit every detail from the Claude conversation. Include the necessary context, acceptance criteria, paths, and constraints in delegated tasks, or use the transfer feature where appropriate.

High-risk changes

For migrations, authentication, authorization, billing, or infrastructure:

Plan → implementation → normal review → fixes → adversarial review → verification

An adversarial review should challenge rollback, race conditions, data loss, trust boundaries, and failure modes.

When one model is stuck

Delegate a focused, evidence-based task:

/codex:rescue --fresh find the root cause and propose the smallest safe fix

Background execution

Long reviews and rescue tasks are better suited to background mode:

/codex:rescue --background implement pagination
/codex:status
/codex:result

Monitor background jobs. A task that appears stuck should be inspected or cancelled instead of consuming usage indefinitely.

Optional review gate

/codex:setup --enable-review-gate
/codex:setup --disable-review-gate

The review gate can block Claude's completion when Codex reports findings.

Enable it only when you can monitor the session. It can create long Claude/Codex loops and consume usage limits quickly.

Codex configuration

Codex reads:

User configuration: ~/.codex/config.toml

Trusted project override: .codex/config.toml

Example:

model = "gpt-5.4-mini"
model_reasoning_effort = "high"

Override a rescue task when needed:

/codex:rescue --model spark --effort medium fix the lint failure

When not to delegate

Keep the task in the current agent when:

The necessary context exists only in the current conversation and cannot be summarized safely.

The task requires continuous product or design decisions from the user.

The task is so small that delegation overhead exceeds the work.

The repository or permission boundary does not allow the delegated operation.

FAQ

The `/codex` commands do not appear

Run /reload-plugins, then /codex:setup. Confirm that the plugin is installed as codex@openai-codex.

A review reports an authentication error

Run /codex:setup, verify Codex CLI authentication, and review Codex CLI Quick Start.

How is `/codex:rescue` different from asking Claude to code?

It delegates the task to the local Codex runtime, which uses its own model configuration and usage limits. The delegated prompt must include the context Codex needs.

Claude and Codex disagree

Treat both outputs as evidence, not authority. Reproduce the concern, inspect the relevant code, and make the final engineering decision.

How much usage does this save?

It depends on task mix, model selection, and pricing. Review actual usage records for both providers.

Claude Code and Codex Collaboration Guide

Install the Codex plugin#

Core commands#

Read-only reviews#

Delegate work#

Manage background jobs#

Transfer the current Claude Code session#

Recommended workflow#

Normal development#

High-risk changes#

When one model is stuck#

Background execution#

Optional review gate#

Codex configuration#

When not to delegate#

FAQ#

The /codex commands do not appear#

A review reports an authentication error#

How is /codex:rescue different from asking Claude to code?#

Claude and Codex disagree#

How much usage does this save?#