Hey Product Hunt — I'm Vince, maker of Model Boss. I built it after noticing that the costly part of AI coding is often keeping the strongest model on the main loop for every implementation detail. Model Boss separates execution from authority: Lite keeps planning and review in the host session, while Max uses a distinct verified reviewer for the plan and final evidence. Workers run in disposable worktrees with OS sandboxing and restricted tools; approvals are bound to the source snapshot, worker delta, and projected patch, so changed evidence invalidates an old approval. The repository includes fixtures, a blind-test probe, and the full benchmark methodology. In its recorded large constructive run, Max reduced strongest-model output tokens by 89% and a price-weighted quota proxy by 88%; these are scoped historical measurements, not a promise for every model or task. It is MIT-licensed and works with Claude Code and Codex. I would love feedback on where you would most want an independent review gate: migrations, refactors, or multi-file features?

Model Boss

Delegate code to cheaper models. Keep frontier judgment.

Delegate code to cheaper models. Keep frontier judgment.