MartinLoop is an open-source governance and orchestration harness designed to manage autonomous AI coding workflows. Rather than acting as a standalone AI coding agent or IDE, it wraps existing agents and models with runtime guardrails, strict budget limits, automated verification gates, and auditable run receipts. In daily operations, developers execute runs through a terminal CLI command (such as `martin run`) or integrate it into IDEs and CI/CD pipelines. Users define tasks, timeboxes, strict spend caps, and verification commands (for example, pointing to a test runner like `pnpm test`). MartinLoop then orchestrates the underlying model or agent—including Claude, Codex, Gemini, OpenHands, or open-source LLMs—evaluating policies, monitoring step-by-step actions, and recording failure taxonomies and telemetry in structured JSONL output. What distinguishes MartinLoop from typical agent frameworks is its explicit focus on cost governance and verifiable completion. Unbounded retry loops are halted when budget ceilings are reached or when repetitive failures occur. An agent cannot declare a task finished without satisfying an independent verifier gate and generating an immutable run receipt detailing commits, diffs, reversibility data, and exact spend. MartinLoop's core CLI and orchestration layer is licensed under the Apache License 2.0 and available at no cost via npm. Enterprise extensions such as MartinLoop360 and HeadlessOS offer additional features including OPA-backed policy-as-code evaluation, business context compilation, and team dashboards.
Tags: agent harness, cli, coding agent, developer tools, OpenSource