About — Undercut

A routing policy, not a company you need to trust blindly.

Undercut is a free, MIT-licensed SKILL.md that coding agents (Claude Code, Codex, Cursor, Copilot, OpenCode) follow at dispatch time. A six-flag rubric assigns each unit of work a base model tier; three objective triggers — a failed verification, a measured disagreement between two cheap-tier attempts, or an explicit uncertainty flag — are the only things allowed to escalate it. Nothing escalates on a vibe or a self-reported confidence score.

It is not a proxy and does not sit on the network path between your agent and its model calls — it's a policy file your agent reads locally, composable with whatever gateway or compression layer you already run.

What's measured, and how to check it

Every cost and pass-rate claim on the landing page traces to a public, reproducible benchmark harness (evals/ in the source repo), run against official test cases with deterministic graders — no LLM judge, no cherry-picked runs. See testing/README.md for the full methodology and raw results, or run it yourself for free with node src/main.js --mock.

Who builds it

Justin Winter — founder and builder. Contact links, source, and license are in humans.txt.

What Undercut is not

• Not a multi-provider gateway or compression proxy — it composes with those instead of competing with them.
• Not a paid product with a free tier held back to make the paid tier look better — the rubric, the eval harness, and every benchmark are identical whether you pay us or not.
• Not venture-scale infrastructure with an API to trust on faith — it's a single auditable file, MIT licensed, forkable, with the numbers behind it public.

Business model

Free for individuals, forever. Teams (org-wide policy enforcement, savings metering, an escalation ledger, SSO) is a paid layer above the free skill, currently in development — see pricing. Enterprise is contact-only.

Updated 2026-08-20. Back to Undercut · Privacy · Terms