The run defines its own stopping point.
- Cost appears after provider calls
- Retries continue without a hard routed cap
- Candidate changes may include the verifier
- Green CI may rely on changed evidence
Public product · open source
Runcap gives AI coding work a cost boundary and an external judge.
The product addresses two different control failures: routed AI calls can spend without a hard ceiling, and AI-generated pull requests can modify the same checks used to approve them.
MIT licensedLocal-first CLIPublic testsGitHub demo PRs
Failure model
Green output is not enough when the actor can influence the judge.
An agent can keep retrying after useful progress stops. A pull request can also alter workflows, tests, or verifiers. Runcap separates spend control from merge evidence and states the scope of each.
Policy is set outside the candidate change.
- Routed work receives a configured ceiling
- A request is blocked before crossing that ceiling
- Proof Gate reads policy from the base commit
- Clean CI emits PASS, BLOCKED, or HUMAN_APPROVAL_REQUIRED
Spend control
Applies only to OpenAI or Anthropic requests routed through the local Runcap gateway. It does not control subscriptions, direct calls, cards, or unrelated tools.
Proof Gate
Current documented scope is GitHub Actions and Node/npm repositories. It checks permitted changes and replays base-pinned verification.
Token evidence
One measured OpenAI edited-file reread fell from 1,186 to 737 prompt tokens, 37.9% fewer, while preserving the correct answer. This is one result, not a universal saving.
Inspect and reproduce
The repository includes installation, current product status, tests, trust model, consumer setup, and token evidence.
Three Proof Gate verdicts
Current product
Read the full feature, setup, and truth boundary.
The product page contains the live demo, installation path, and current limitations.