OpenAI Codex + Claude Code Guide

OpenAI Codex and Anthropic’s Claude Code are both agentic coding tools. Each one can read a codebase, edit files, run commands, and review changes.

Running both gives you:

• A second opinion on high-stakes changes
• Independent reviews from separate model families
• A fallback when one tool hits its usage limit
• Access to each tool’s ecosystem strengths
• Parallel progress on separate tasks

Two agents are not automatically twice as good. They pay off when each one has a clear role.

Made by

OpenAI

Surfaces

ChatGPT desktop app, CLI, IDE extension, Codex Cloud, GitHub, Slack, and mobile

Recommended models

GPT-6.1 Sol for complex work, GPT-6 Luna for focused tasks, and GPT-6 Astra for the toughest problems

Instructions file

AGENTS.md

Skills folder

.agents/skills

Plans

Included from ChatGPT Plus ($20/month); Free and Go users get GPT-6 Luna in the desktop app, depending on rollout

Made by

Anthropic

Surfaces

Terminal, VS Code, JetBrains, Claude desktop app, web, mobile, and Slack

Models

Claude Opus 5.5, Sonnet, Haiku, and more, with effort levels from low up to max

Instructions file

CLAUDE.md, and it can also read AGENTS.md

Skills folder

.claude/skills

Plans

Part of every paid Claude plan, starting with Pro at $20/month

Both also offer Windows installers, Homebrew, and npm. Run codex and claude inside your project folder to sign in.

Keep a single source of truth so both agents work from the same rules.

• Put your project rules in AGENTS.md
• Let Codex read AGENTS.md automatically
• Let Claude Code read AGENTS.md when no CLAUDE.md exists (v2.1.277 or later)
• Add @AGENTS.md to CLAUDE.md to import the shared file if you need both

One rulebook, two agents. Never keep two copies of the same rules.

Cover these in AGENTS.md:

• What the project does
• Setup and install commands
• Test, lint, and build commands
• Code style and naming conventions
• Folders and files to leave alone
• Actions that require human approval
• Commit message format
• What counts as done

Keep it short and specific. By default, Codex caps combined AGENTS.md content at 32 KiB.

OpenAI publishes an official open-source plugin that runs Codex from within Claude Code. It calls the Codex CLI already on your machine, using your current login and settings.

Install it from Claude Code:

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

You need a ChatGPT subscription (Free included) or an OpenAI API key, plus Node.js 18.18 or later.

/codex:review

Run a standard read-only review of your current changes

/codex:adversarial-review

Run a review that challenges design choices and assumptions

/codex:rescue

Hand Codex a bug investigation, fix, or follow-up task

/codex:transfer

Convert the current Claude Code session into a Codex thread

/codex:status

List running and recent Codex jobs

/codex:result

Display the output of a finished job

/codex:cancel

Stop a background job mid-run

The simplest high-value pattern: one agent builds, the other reviews.

• Claude Code makes the change
• You run /codex:review or /codex:adversarial-review
• Claude Code fixes the valid findings
• You check the final diff and tests

Or flip it: Codex builds, and Claude Code reviews the diff.

Treat every review as evidence to verify, not a verdict to accept.

For bigger features:

• Have one agent draft a detailed plan
• Ask the other to poke holes in it
• Merge the best ideas yourself
• Have one agent build from the final plan
• Have the other review the result

Disagreement between the agents helps. It points you to the real decisions.

Never let two agents edit the same files at once.

• Put each agent on its own Git branch
• Use separate Git worktrees for parallel work
• Assign tasks that do not overlap
• Merge changes one at a time
• Run tests after each merge

In Codex, /worktree runs a chat in a fresh Git worktree.

Codex Cloud

Run several tightly scoped tasks remotely, then review what comes back.

Codex GitHub review

Mention @codex review on a pull request or switch on automatic reviews.

Claude Code skills and hooks

Package workflows and enforce checks before or after each action.

Claude Code routines

Run scheduled jobs in the cloud, even while your computer is off.

Both

MCP connections, subagents, and terminal-based workflows.

Test both tools on your own codebase. Real tasks tell you more than benchmarks.

Both tools define skills in a SKILL.md file with a name and a description.

• Codex loads skills from .agents/skills
• Claude Code loads skills from .claude/skills
• Shared logic belongs in scripts both can run
• Tool-specific notes stay short

Check each tool’s docs for format differences before you copy a skill across.

Both support the Model Context Protocol, so you can wire the same docs, issue tracker, or database into each one.

Codex

codex mcp add

Claude Code

claude mcp add

Give each agent only the servers its role requires.

Make both agents ask before they push, deploy, delete, or touch secrets.

• Pick the cheapest model that does the job well
• Raise effort only for genuinely hard problems
• Save reviews for important changes, not every typo
• Switch tools when one hits its limit
• Keep instructions lean to save context

In Codex, /model changes both the model and reasoning effort. In Claude Code, use /model and /effort.

Review these yourself:

• Architecture decisions
• Security-sensitive code
• Database migrations
• Payments and authentication
• Anything that ships to production
• Disagreements between the agents

Both companies warn that their agents can make mistakes. Two agents agreeing proves nothing.

Rules

A single shared AGENTS.md.

Plan

One agent plans and the other critiques.

Build

One agent builds on its own branch.

Review

The other agent checks the diff.

Fix

Resolve the valid findings.

Test

Run the whole test suite.

Approve

A human reviews, then merges.

Learn

Add new lessons to AGENTS.md.

Hand this to the agent doing the build:

Read AGENTS.md and follow it exactly. Goal: [describe the change]. Constraints: [performance, security, or style requirements]. First, write a short plan and list the files you expect to touch. Wait for my approval. Then make the change on a new branch, add or update tests, and run the full test suite. When you finish, summarize what changed, why, and what a reviewer should check most closely. Do not push, deploy, or delete anything without asking me.

Then have the other agent review that summary and the diff, with a focus on correctness, security, and missing tests.