MartinLoop
An open-source AI coding agent runtime that works with Claude Code, Codex CLI and Gemini CLI, with local run records and enforced budgets.
MartinLoop is a free, self-hosted runtime for developers and engineering teams delegating software work to AI coding agents. It wraps Claude Code, Codex CLI and Gemini CLI with limits and checks that sit outside the agent. The agent writes the code; MartinLoop decides when work can continue and when the result passes the configured checks.
Runs have spending, time and attempt limits, plus boundaries on which paths an agent can change. MartinLoop stops stalled work and requires verification evidence before accepting completion. Each run ends with a verified result, a stop reason or a request for review. Recovery and rollback records support another attempt or a handoff, while receipts bring together checks, failures and costs.
The job definition stays separate from the coding agent, so teams can switch workers while retaining the same limits and verification rules. MCP integration connects it to compatible hosts. OpenAI-compatible inference workflows can use hosted providers or local endpoints such as Ollama, LM Studio and llama.cpp; that path isn't presented as a finished autonomous coding worker.
The core uses the Apache 2.0 license and doesn't require an account to start. Run records stay local and carry signatures that make later edits detectable. Anonymous usage telemetry is enabled by default after disclosure and can be disabled. It excludes source code, prompts, repository contents, model output and receipt contents. A separate hosted dashboard is also offered.
