Code Duo

Code Duo

Stop trusting one AI. Put two in the same room and make them check each other.

Claude and Codex, side by side in one window. Route a prompt to either or both, watch each one work step by step, hand a half-finished change from one to the other, and keep a watchdog on whether the AI actually did what it claimed — or just burned your tokens saying it would.

No API keys, ever. Code Duo drives the claude and codex CLIs you already pay for, on your existing Max / ChatGPT subscriptions.

The problem this solves

You give an AI a real task. It says "no problem, I've got this," sounds confident the whole way, burns a pile of tokens — and at the end shrugs and tells you it "overestimated the difficulty." The work never lands. You can't see it going in circles until it's already cost you an hour.

One confident AI will talk you into a wall. Two AIs, watching each other, won't.

What you get

- Two agents, one window, zero context-switching. Press Tab to aim at Claude, Codex, or Both. Both runs the same prompt through both in parallel so you can compare two solutions instantly.
- Watch the work, not just the answer. Live streaming of every step — reasoning, commands run, files edited, reads — so you catch a bad path early instead of after it's burned your budget.
- Cross-review and hand-off, built in. One click sends one agent's work to the other to take over or audit, all on the same project files. When Claude is stuck, Codex gets a fresh crack at the exact same problem — on screen, side by side.
- A watchdog for AI bullshit. Every turn compares what the AI *claimed* against what *actually changed on disk*. It flags busywork ("5 actions claimed, 0 files changed") and going in circles ("looping 3×").
- Your real sessions, not a blank slate. Your actual Claude and Codex…