Describe a feature. Come back to a shipped, proven change.

How it works

Three moves. You make the first one; the factory makes the other two.

1. 1

   Describe the change

   A brief the way you would tell a colleague.

2. 2

   The factory builds and proves it

   Claude Code works in your repository on a workstation. Then the gate runs: lint, types, tests, a build, Playwright with video.

3. 3

   It lands, with evidence

   A script rebases the candidate and pushes it to your default branch. You get the diff, the video, the agent’s report and the cost.

What keeps it honest

The four doubts you should have, answered by how the factory is built.

Your repository, your rules.

The unmodified Claude Code runs with your CLAUDE.md, skills and hooks, on a workstation the factory provisions, never on your laptop. The repository, the database and the deploy stay yours.

Nothing lands without proof.

Every run passes the gate you would run yourself and records it: results, Playwright video and traces, screenshots, the diff, the agent’s report, cost. The lander is a script. No model touches your default branch.

It asks, then waits.

When the agent needs a decision, the run pauses and the question reaches your phone. Answer, and it resumes where it stopped. New work that changes schema, money, auth or policy waits for your explicit approval; a fix lands by itself.

Your subscription, your credential.

Bring the token from your own Claude subscription, or an API key. It is sealed, delivered only to the project’s workstation for a run, and never shown again.