I ran the Codex ↔ Claude Code review loop by hand for months, so I built a canvas where it runs itself

Codex is my main driver. On bigger tasks I add a second model—Claude Code reviewing Codex's diffs, or Codex + Claude + Grok debating a design. But the loop itself was me: paste the diff to the reviewer, carry blocking issues back, decide when it's clean enough to stop.

So I built looperators. The loop now runs on a canvas: Codex finishes and the diff goes to the reviewer automatically; blocking issues come back and wake the same Codex session to fix them; it stops at "review clean, max 6 laps." A badge on the ring shows the lap count, and every agent stays a real session you can open mid-run and talk to.

Everything runs on the CLIs you already have (Codex, Claude Code, Grok Build), orchestrated locally—your existing Codex login, no API keys proxied, no model resale.

try it:

https://github.com/ObservedObserver/looperators