What is
AgentMachinist?

AgentMachinist helps you turn an idea into a code change you can check before accepting it.

You approve the plan, your AI coding assistant does the work, and AgentMachinist organizes the tests and a separate review. You decide what becomes part of your project.

From an idea to an accepted change.

AgentMachinist manages the steps. You make the two decisions.

  1. You

    Describe a Task

    Name one small change you want to make.

  2. Your assistant

    Write the plan

    Explain the change and how to check it. AgentMachinist saves this plan.

  3. Your decision

    Approve the plan

    Read it first. Work waits for your Approval of that exact plan.

  4. Your assistant

    Make the change

    Edit a separate working copy of your project.

  5. AgentMachinist

    Check and review

    Run your project's checks and arrange a separate Review of the change.

  6. Your decision

    Accept the change

    Inspect the result. You choose whether to add it to your project.

Orange marks your steps. Blue marks the machine's work. After Review, you can also choose to share the result through a GitHub pull request or GitLab merge request.

A change with the evidence beside it.

Before accepting, you can read what was planned, what changed, and what the checks and Review found.

  • The plan you approvedWhat the change was meant to do.
  • The exact code changesWhat your assistant actually changed.
  • Your project's check resultsWhich checks passed or failed.
  • A separate ReviewFindings to help you inspect the work.

Works with the assistant you already use.

AgentMachinist runs on your computer and works with Claude Code, Codex, OpenCode, Pi, or Goose. Your assistant writes the code; AgentMachinist organizes the workflow and keeps the record.

The core workflow runs in your local project. GitHub and GitLab are optional. Your assistant can still use a cloud model, and real Tasks use your assistant's plan or API credits.

The checks and Review help you assess the change. Review findings are advisory, and you still make the final decision. Read the trust model.

Try one small change.

Start with the free rehearsal. It uses a fake assistant to walk through the workflow, with no model calls or API cost.

Try it free Watch the 90-second tour