One writer
Only Codex edits the isolated worktree. The reviewer cannot race the implementation or silently overwrite the same files.
OPEN-SOURCE · CODEX + CLAUDE CODE · LOCAL DESKTOP
Duet.
Codex builds. Claude challenges. Machine checks decide.
Duet is a local, single-writer review loop that runs Codex in an isolated worktree, brings Claude Code in as a fresh read-only reviewer, and lets repository checks govern completion.
HOW IT WORKS
A bounded path from request to evidence.
- 01
Isolate
Create a dedicated Git worktree and fingerprint the starting repository state before either agent begins.
- 02
Build
Codex remains the only writer, implementing the task with the repository and its declared constraints in scope.
- 03
Challenge
Claude Code reviews from a fresh, read-only context and returns compact findings rather than sharing or inheriting the writer transcript.
- 04
Apply
Machine checks decide readiness; guarded apply and exact-state undo keep the developer in control of the main workspace.
DESIGN
The product is the contract.
The interface is useful because it exposes what happened, where confidence comes from, and where the system stops—not because an agent produced a confident paragraph.
Fresh critical context
Claude Code receives the repository state and a bounded handoff, not the implementation transcript, preserving a genuinely independent review pass.
Machine checks win
Neither agent can declare success over a failing repository command. The configured check result is the completion gate.
Finite rounds
The loop is explicitly bounded from one to six review rounds so disagreement cannot turn into an unmetered conversation.
Exact-state recovery
Repository fingerprints and guarded undo detect intervening changes before restoring the pre-apply state.
Local, transcript-free receipts
Duet records compact operational evidence locally without turning private agent transcripts into a hosted collaboration service.
QUICK START
Inspect the real thing.
This is the current source path. Duet expects the official Codex and Claude Code CLIs to be installed and already signed in; signed and notarized desktop installers are not published yet.
TERMINALSOURCE-READY ALPHA · v0.1.1 CANDIDATE
$ git clone https://github.com/chriswu727/agent-duet.git
$ cd agent-duet
$ corepack enable && pnpm install --frozen-lockfile
$ pnpm startVERIFIABLE SCOPE
- 99
- offline tests
- 3
- runner platforms
- 1–6
- bounded review rounds
- ZERO
- API-key fallback
The offline suite validates orchestration and safety contracts. The v0.1.1 candidate has not yet cleared signed-installer distribution, external beta, or a live subscription smoke test.
HONEST BOUNDARIES
What this does not prove.
- 01Duet is currently a source-ready alpha; signed and notarized desktop installers have not been published.
- 02The v0.1.1 package is a candidate rather than a completed public release, and the external beta gate remains open.
- 03Live Codex and Claude Code subscription smoke testing has not been claimed here; it consumes the user's existing plan limits and requires a real signed-in environment.
- 04Duet deliberately has no API-key or usage-billing fallback. It depends on supported behavior from the official subscription CLIs.
- 05It is a local single-user safety loop, not a hosted multi-user proxy or a promise that two models will agree on every task.
FAQ
Before you try it.
Does Duet require separate API credits?
No. It is designed around the official Codex and Claude Code subscription CLIs and intentionally does not fall back to API keys. Normal subscription plan limits still apply.
Can both agents edit the repository?
No. Codex is the sole writer inside an isolated worktree. Claude Code reviews from a fresh read-only context and returns findings for the writer to address.
Are signed installers available?
Not yet. The current v0.1.1 candidate is source-ready alpha software. Signed and notarized distribution remains a release gate, so the product page points to source instructions and the readiness document.
Why limit the number of review rounds?
A finite one-to-six-round budget bounds time and subscription usage, makes non-convergence explicit, and prevents an open-ended agent conversation from masquerading as progress.
OPEN SOURCE