Claude Code writes it. Codex reviews it. Duo runs both

Duo is an MIT-licensed local desktop app and CLI that runs your signed-in Claude Code and Codex CLIs together.

Its pair mode is more useful than asking two models for competing answers, because each model gets a different job:

• One model writes the code in an isolated git worktree.
• The second model reviews the diff against the request in read-only mode.
• Duo runs a check command, such as your test suite.
• The writer fixes or disputes the findings, and the loop repeats.

It stops when the writer reports completion, the reviewer approves without serious findings, and the check passes. Nothing touches your original project folder until you choose Apply, Keep Branch, or Discard.

Prompts, replies, tool calls, checks, usage, reasoning traces, and run artifacts are all recorded locally. That makes every run traceable.

Installation adds a Duo skill for Claude Code, a Duo skill for Codex, and a Codex rule that pre-approves a restricted duo-safe entry point. Inspect those files and understand their permissions before relying on them.

For a first test, use a disposable repository and one small bug. Keep both agents sandboxed, inspect the worktree and the trace, then discard the result instead of applying it. Trust the process only after you have tested its isolation and cleanup.

The catch: running two models can use up substantially more of your plan.