Building MartinLoop — an open-source control plane for AI coding agents.

The problem: AI coding agents (Claude Code, Codex, OpenCode) have no built-in brakes. They retry indefinitely, rack up cost, and give you nothing auditable at the end. I had a run estimated at $2.40 hit $65 before I caught it. MartinLoop wraps any agent with: Hard budget caps (USD, tokens, or iterations) Verifier gates — npm test must pass before a run counts Full audit trail — every attempt, cost, and outcome in JSONL Run receipts — one command shows exactly what happened and why it stopped It's open-source, TypeScript, works in 5 minutes alongside your existing setup. Would love feedback from anyone building with AI coding agents — especially around how you currently handle cost control or verification. martinloop was originally created to solve the agent loop and ralp loop problems. What's broken for you? for me its runaway agents working with infinite loops its constant task drift on large context work agent teams being less productive than single agents consistent and reliable output from coding agents that is PRD ready never understanding what is actually happening with claude, needed a receipt and brakes. MY ask: would love any feedback, appreciate an upvote🙏🏽 if you like it, and please help me out by starring ⭐ on github to keep the core open source.