Demo · A governed coding-agent job

    See a coding-agent job run with boundaries, checks, and a clear outcome.

    The published flaky-CI-gate benchmark is one task-specific example: $2.30 with MartinLoop versus $5.20 ungoverned, with one attempt versus four failed retries.

    Setup

    One npm install. That's the whole setup.

    20-second install · zero configuration · nothing to migrate to

    Benchmark snapshot: same task, same start, different governance
    Head-to-head comparison

    Same task. Different job contract.

    ralph-loop (traditional)
    $ agent run --task "flaky-CI-gate"
    [retry 1] ✗ Verification did not pass
    [retry 2] ✗ Verification did not pass
    [retry 3] ✗ Verification did not pass
    [retry 4] ✗ Verification did not pass
    Total cost: $5.20
    Failed retries: 4
    Result: Not completed in the published comparison
    $5.20 ungoverned · 4 failed retries
    Animated MartinLoop CLI run: distilled context, taxonomy routing, evidence-gated completion
    $2.30 MartinLoop · 1 attempt

    Stops at your budget, every time

    A hard ceiling per loop. The agent doesn't get to argue with it.

    Fixes the right thing, not everything

    Knows the difference between a lint error and a hallucination — and fixes each one differently.

    Replayable record of every run

    A full timeline you can hand to engineering, finance, or audit. No screenshots required.

    Powered by the 13-class failure taxonomy in @martin/contracts.

    Real run · 60-second highlight

    A real MartinLoop run: same task, evidence-gated exit, receipt attached.

    Watching MartinLoop govern a real gradient bug fix — from plan through evidence-gated completion.

    First 60 seconds of a 5-minute governed run.

    Proof of work

    Every run opens with a plan and closes with a receipt.

    The plan and the final run record make the job inspectable before and after execution.

    MartinLoop governed run plan captured before execution
    Before the runRun plan captured before execution
    MartinLoop final run receipt showing the recorded outcome and evidence
    After the runRun receipt captured after execution

    Published benchmark snapshot.

    One reproducible task, not a universal savings claim.

    Cost
    $5.20 ungoverned
    $2.30 MartinLoop
    task-specific
    Attempts
    4 failed retries
    1 attempt
    published comparison
    Hosted dashboard

    See the run after the agent stops.

    Inspect the outcome, verified receipt evidence, and trusted spend for the workspace. Hosted dashboard and team workflows are in early access.

    MartinLoop hosted dashboard showing run outcomes, receipt evidence, and trusted spend

    Free forever on the open-source tier. No credit card needed. We only store your email and the tier you’re interested in — see our Privacy Policy. Request deletion anytime at help@martinloop.com.

    See it live

    What the hosted dashboard makes inspectable.

    Outcome state, verification evidence, stop reason, and trusted spend stay visible after the agent stops.

    Verified

    Configured checks passed

    Verification evidence and receipt integrity remain inspectable.

    Stopped

    A boundary fired

    The stop reason and run record remain available for review.

    Needs review

    Evidence was not enough

    The run stays inspectable instead of being presented as complete.

    Sign in required. Trusted spend is counted only when supported by the hosted evidence rules for that workspace.

    Give your AI coding agents a finish line.

    Every blind retry doubles the bill. Every "almost done" wastes another loop.

    MartinLoop gives the loop a finish line. The test has to actually pass before the run is called done. The budget can't be negotiated. The history is replayable.

    Install it. Run one task. See exactly what happened.

    $npx -y martin-loop@latest start

    Free forever on the open-source tier. No credit card needed. We only store your email and the tier you’re interested in — see our Privacy Policy. Request deletion anytime at help@martinloop.com.