How to use Claude Code and Codex together on one repo

Why run both

A second opinion. A model reviewing its own work tends to approve it. A different vendor's model catches different mistakes.
Best of two. Give both agents the same task and keep the better branch. On hard bugs and open-ended features the two attempts often differ a lot.
Different strengths. Many people prefer one agent for planning and refactors and the other for tight, well-specified changes. Running both lets you route each task to the one you trust for it.
Two sets of usage limits. Claude Code uses your Claude plan or API key and Codex uses your ChatGPT plan or OpenAI API key, so when one hits its limit the other can keep working.

Step 1: one instructions file for both agents

Codex reads project instructions from AGENTS.md. Claude Code reads CLAUDE.md, and since v2.1.277 it also reads AGENTS.md when the repo has no CLAUDE.md (Anthropic's docs, October 2026). So the simplest setup is a single AGENTS.md at the repo root with your build commands, test commands, and conventions.

If you already have a CLAUDE.md, keep it and import the shared file from it, so both agents get the same rules:

# CLAUDE.md
@AGENTS.md

# Claude-only notes go below the import.

Step 2: a worktree per agent

Two agents in one folder edit the same files on the same branch, and their changes get mixed together. Give each its own git worktree: a separate checkout of the same repo on its own branch.

cd ~/code/myapp
git worktree add ../myapp-claude -b search-claude
git worktree add ../myapp-codex  -b search-codex

Each worktree only has tracked files, so install dependencies and copy untracked config such as .env into each one. Then start one agent per worktree, in two terminals:

cd ~/code/myapp-claude && claude
cd ~/code/myapp-codex  && codex

The details of running each agent in parallel (tmux, ports, resuming, merging) are in the guides for Claude Code and Codex.

Step 3: race them on the same task

Paste the same prompt into both sessions. When both are done and have committed, compare the two attempts from the main checkout:

cd ~/code/myapp
git diff --stat main search-claude
git diff --stat main search-codex
git diff search-claude search-codex   # where the two attempts differ

Run the tests in each worktree, keep the better branch, and delete the other.

Step 4: have each agent review the other

Cross-review works the same whether you raced them or split the work. In Claude's worktree, ask Codex for a review of the branch against main:

cd ~/code/myapp-claude
codex review --base main

And in Codex's worktree, ask Claude Code with a one-off, non-interactive prompt:

cd ~/code/myapp-codex
claude -p "Review the changes on this branch against main (git diff main...HEAD). List bugs and risky changes, most serious first. Don't edit files."

Paste the findings back into the session that wrote the code and let it fix them. Then merge or open a pull request as usual, and clean up with git worktree remove.

Gotchas

Two accounts. Each CLI signs in on its own. Claude Code needs a Claude Pro, Max, Team, Enterprise, or Console account (the free Claude plan doesn't include it). Codex uses a ChatGPT plan or an OpenAI API key.
Agent-specific config stays separate. MCP servers, permissions, and hooks are configured per tool (.claude/ for Claude Code, ~/.codex/config.toml for Codex), so set up anything both need in both places.
Ports and services. Worktrees isolate files, not ports or your local database. Give each worktree its own dev server port, and a separate database if the agents run migrations.
One branch per worktree. Git won't check out a branch that's already checked out in another worktree, which keeps the two agents off each other's branch.

The easier way: Maestro

Maestro is a free desktop app built for this. You choose the agent per workspace, and every workspace is its own git worktree and branch, so Claude Code and Codex (plus Cursor, OpenCode, Kimi Code, and Grok Build) run side by side in one window, each with its chat, terminal, and diff. Each agent uses its own CLI and your own account, as it does in the terminal.

Review either agent's diff, comment on lines, and send the comments back as its next message.
Open a pull request for any workspace from the app. There's no limit on workspaces.
Follow every agent from your phone with Maestro Web.
macOS, Windows, and Linux.

Download MaestroFree · macOS, Windows & Linux

More on each agent in Maestro: Claude Code GUI and Codex GUI.

FAQ

Can Claude Code and Codex work in the same folder at once?

They can, but they'll edit the same files on the same branch and you'll get one mixed diff. Give each its own worktree, as above.

Do I need to keep both CLAUDE.md and AGENTS.md?

No. On current Claude Code versions an AGENTS.md alone works for both agents. Keep a CLAUDE.md only for Claude-specific notes, and import @AGENTS.md from it.

Which one should I use for what?

Models change every few months, so test both on your own codebase: race them on a few real tasks (step 3) and keep notes on which one wins which kind of task.

Maestro is an independent project, not affiliated with Anthropic or OpenAI. Claude Code and Codex details above come from their public docs as of October 2026.