Martin Loop

MartinLoop wraps AI coding agents like Claude Code with hard budget caps, verifier gates, and structured audit trails. Each run gets a contract: max spend in USD or tokens, a verifier command like `npm test`, allowed file paths, and a stop condition that isn't just "the agent gave up." The MCP server exposes `martin_run` for governed execution plus read-only tools to inspect run records, attempt history, verifier results, and budget consumption. Reach for this when you need to cap retry loops before they burn through tokens, require proof that tests passed before marking work complete, or build an audit trail showing exactly what changed, what it cost, and why it stopped.