This website uses cookies
Read our Privacy policy and Terms of use for more information.
The setup I use for every build: Claude Code writes the change, and a gate decides whether it is ready to merge. Three steps.
Use the terminal version if you can. It runs next to git and your editor, and it is the one I use daily. The web app (claude.ai/code) and the desktop app (claude.com/download) work too. Follow Anthropic's own setup guide; install steps change, and that page stays current.
Put a CLAUDE.md file in the project root: how to run the tests, what must never change, and what done means. The agent reads it at the start of every session. I walk through mine in The file that tells Claude how you work.
An agent will tell you its work is finished. That is not evidence. I use Keel, which I built for this. For a fix or a feature, a new test must fail on the old code and pass on the new code. The tests and rules are locked, so the agent cannot weaken them. A model from a different family reviews the diff. Then you read a PR that carries the evidence.
It needs Node 24 or later, git, and a logged-in claude or codex CLI. It is not on npm yet, so install it from the repo:
Keel is not a sandbox. The agent runs with your user's rights, so keep production secrets out of the environment.
Looking for Claude Kit? It is retired. I archived it on 27 September 2026 and moved to Keel. See what is current on Builds.
One build like this every Thursday, with what it cost, what broke, and the evidence I checked. Subscribe free →