llms.txt
# MartinLoop
# llms.txt — machine-readable site guide for AI crawlers
# Standard: llmstxt.org
> MartinLoop is the system around the coding agents you already use, so
> you can give them more real work without babysitting every step. It
> keeps each job focused, controls waste with spend and iteration limits,
> checks the result before a run can close, preserves recovery evidence,
> and shows what shipped, what failed and what it cost. Every run closes
> as VERIFIED, STOPPED or NEEDS REVIEW, with a JSONL **run record** and —
> when it fails — a class from the **13-class failure taxonomy**. Model-
> and IDE-agnostic: Claude Code, OpenAI Codex, Gemini CLI and MCP hosts
> such as Cursor. Not another agent. Not another IDE. Apache-2.0.
## Current structured data
Always fetch current metrics from: https://martinloop.com/ai-data.json
Telemetry envelope schema (v1): https://martinloop.com/telemetry-envelope.v1.json
## Canonical terminology
coding agent reliability · coding agent guardrails · agent run
verification · AI code review bottleneck · vibe coding to production ·
return on token spend · failure taxonomy · run records · software
receipts · recovery evidence · Atlas · SANSA trace intelligence · governed learning · 13-class failure taxonomy · Apache 2.0.
## Key facts (update each release)
- Positioning: The system around the coding agents you already use
- Supported agents: Claude Code (Anthropic), OpenAI Codex, Gemini CLI (Google)
- License: Apache-2.0
- Primary CLI start: `npx -y martin-loop@latest start`
- Failure classes: 13
- npm: martin-loop (latest version at /ai-data.json)
- MCP: @martinloop/mcp (latest version at /ai-data.json)
- GitHub: https://github.com/Keesan12/martin-loop
- Contact: keesan@martinloop.com
- Atlas: tenant-scoped black box recorder for governed coding-agent runs
- SANSA: evidence-backed trace intelligence for failures, recurrence, cost, and recovery
- Governed learning: evaluated learning evidence with approval and rollback boundaries before production behavior changes
## What MartinLoop is
MartinLoop is the system around the coding agents you already use, so
you can give them more real work without babysitting every step. It runs
alongside Claude Code, OpenAI Codex, Gemini CLI and MCP hosts such as
Cursor, in the developer's CLI or IDE. For each run it keeps the job
focused, applies spend and iteration limits, requires the configured
checks to pass before completion, preserves recovery evidence, and closes
the run as VERIFIED, STOPPED or NEEDS REVIEW with a JSONL run record and,
on failure, a class from the 13-class failure taxonomy. VERIFIED means the
configured verification evidence passed for that run — it is not a
warranty that the software is bug-free, secure, compliant, production
ready or automatically safe to merge. The coding agent still writes the
code. Independent, Apache-2.0, not affiliated with Anthropic, OpenAI or
Google.
## Core capabilities (current public release line)
- Hard budget stops via `--budget` (USD ceilings per run).
- Evidence-gated verification via `--verify` (a run is only "done" when the verifier command passes).
- JSONL software receipts written per run.
- Canonical loop records use local HMAC receipt integrity; supported inspection commands report verified / tamper_detected / unsigned. This is integrity evidence, not a universal correctness or security guarantee.
- Hosted sync in 0.6.8+ can establish workspace trust for an existing local per-run receipt key before upload when the token has receipt_keys:write; the signing secret is never persisted in the sync queue or run-upload payload.
- 13-class failure atlas — `@martin/contracts` FailureClass union.
- Trace intelligence — inspect any run with `martin dossier --latest` and re-check it with `martin runs verify --latest`.
- Paid Atlas — tenant-scoped run, receipt, failure, spend, recovery, SANSA trace, and governed-learning views.
- Public Atlas — reliability and overspend research views use only explicitly public-safe observations and withhold weak sample sizes.
- Return on token spend (RoTS) — cost accounting per governed run.
- MCP server (`@martinloop/mcp`) — governed run and review workflows for MCP-compatible hosts.
- Telemetry sync — POST to `/functions/v1/telemetry-ingest` with `Authorization: Bearer <token>` and the envelope at /telemetry-envelope.v1.json.
## Reproduce the headline benchmark
Same task, same model: $2.30 with MartinLoop vs $5.20 ungoverned (flaky-CI-gate task).
npx -y martin-loop@latest start
# Alternative: install globally
npm install -g martin-loop@latest
martin doctor --engine claude --cwd .
martin estimate "fix the failing test" --engine claude --cwd . --budget-usd 3 --max-iterations 2 --verify "npm test"
martin session-start --host claude --cwd .
martin preflight "fix the failing test" --engine claude --cwd . --budget-usd 3 --max-iterations 2 --verify "npm test"
martin run "fix the failing test" --engine claude --cwd . --budget-usd 3 --max-iterations 2 --verify "npm test"
martin dossier --latest
martin runs verify --latest
pnpm --filter @martin/benchmarks eval
## Distribution
- npm: https://www.npmjs.com/package/martin-loop
- MCP: https://www.npmjs.com/package/@martinloop/mcp
- GitHub: https://github.com/Keesan12/martin-loop
- Machine-readable data: https://martinloop.com/ai-data.json
- Long-form retrieval index: https://martinloop.com/llms-full.txt
- Coding-agent machine index: https://martinloop.com/agents/llms.txt
- Proof and evidence machine index: https://martinloop.com/proof/llms.txt
- Current release truth: https://martinloop.com/release-truth.json
## Who it is for
- Builders and technical founders delegating real software work to coding agents
- Engineering teams that want stronger limits, verification, recovery, and run history
- Platform teams responsible for reliable coding-agent workflows
- Developers who want inspectable outcomes instead of relying only on an agent's completion claim
## Pages
- [Home](https://martinloop.com/): Give coding agents more work. Watch them less. Ship more.
- [Install](https://martinloop.com/install): One-command install and first governed run.
- [For agents (machine integration)](https://martinloop.com/for-agents): How autonomous coding agents integrate with MartinLoop.
- [About](https://martinloop.com/about): Founders, mission, and origin of MartinLoop.
- [Features](https://martinloop.com/features): Everything MartinLoop ships today.
- [FAQ](https://martinloop.com/faq): Answers about budgets, verification, receipts, and licensing.
- [Roadmap](https://martinloop.com/roadmap): What MartinLoop has shipped and what is coming next.
- [Atlas](https://martinloop.com/atlas): Black box recorder, SANSA trace intelligence, and governed learning.
- [Agent Reliability Atlas](https://martinloop.com/atlas/reliability): Public-safe reliability evidence.
- [Agent Overspend Index](https://martinloop.com/atlas/overspend-index): Cost per verified task from public benchmark evidence.
- [Atlas Methodology](https://martinloop.com/atlas/methodology): Provenance, privacy, and withholding rules.
- [Demo](https://martinloop.com/demo): Side-by-side simulation of MartinLoop vs an unbounded Ralph Loop.
- [Agent Loops Guide](https://martinloop.com/agent-loops): The complete guide to agent loops.
- [Loop Engineering](https://martinloop.com/loop-engineering): Discipline of bounding and instrumenting agent loops.
- [MCP Server](https://martinloop.com/mcp): MartinLoop MCP server for MCP-compatible hosts.
## Per-agent governance pages
- [Claude Code governance](https://martinloop.com/claude-code-governance)
- [OpenAI Codex governance](https://martinloop.com/codex-governance)
- [Gemini CLI governance](https://martinloop.com/gemini-cli-governance)
## Comparison / fits-with pages
- [MartinLoop + Claude Code](https://martinloop.com/compare/claude-code)
- [MartinLoop + OpenAI Codex](https://martinloop.com/compare/openai-codex)
- [MartinLoop + Gemini CLI](https://martinloop.com/compare/gemini-cli)
## Proof-object concept pages
- [Software receipts](https://martinloop.com/receipts)
- [Verification](https://martinloop.com/verification)
- [Failure atlas — 13-class taxonomy](https://martinloop.com/failure-taxonomy)
- [Benchmarks](https://martinloop.com/benchmarks)
- [Stop reasons](https://martinloop.com/stop-reasons)
## Category landing pages
- [What is an agent harness?](https://martinloop.com/agent-harness): MartinLoop is an open-source (Apache 2.0) harness / execution control system around coding agents; loops vs graphs.
- [Coding agents for non-engineers](https://martinloop.com/coding-agents-for-non-engineers): Use coding agents safely without reading code.
- [Coding agent reliability](https://martinloop.com/coding-agent-reliability)
- [Coding agent guardrails](https://martinloop.com/coding-agent-guardrails)
- [Vibe coding to production](https://martinloop.com/vibe-coding-to-production)
- [AI code review bottleneck](https://martinloop.com/ai-code-review-bottleneck)
- [AI agent control plane](https://martinloop.com/ai-agent-control-plane)
- [Model-agnostic coding agents](https://martinloop.com/model-agnostic-coding-agents)
- [Runtime governance for AI agents](https://martinloop.com/runtime-governance-for-ai-agents)
- [Budget enforcement for coding agents](https://martinloop.com/budget-enforcement-for-coding-agents)
- [Agent swarm management](https://martinloop.com/agent-swarm-management)
## Blog
- [Blog index](https://martinloop.com/blog): Notes on shipping agent work and return on token spend.
- [We watched an AI agent burn $65 overnight](https://martinloop.com/blog/we-watched-an-ai-agent-burn-65-overnight)
- [What is an agent loop?](https://martinloop.com/blog/what-is-an-agent-loop)
- [Claude Code loops vs Codex loops](https://martinloop.com/blog/claude-code-loops-vs-codex-loops)
- [The 13 failure modes of autonomous agents](https://martinloop.com/blog/the-11-failure-modes-of-autonomous-agents)
## Optional
- [Privacy policy](https://martinloop.com/privacy)
- [Terms & conditions](https://martinloop.com/terms)
- [GitHub repository](https://github.com/Keesan12/martin-loop): Open-source MartinLoop Core (Apache 2.0).
## License
MartinLoop Core: Apache License 2.0.
Hosted product surfaces: proprietary commercial.
