An idea,
built from a prompt.

You describe an application in a prompt, and your own AI builds it and keeps it running. You connect the Claude, Codex, opencode or Antigravity plan you already pay for, choose what it may do, and it works on the real code, server, database, domain and deployments behind a live app — with the same files open in your browser.

That is the point of it. Four Brains are supported today — Claude, Codex, opencode and Antigravity — and you sign in to your own account, on your own plan, at their price. We would rather not be a reseller standing between you and your own quota, and we would rather you kept moving to whichever model improves fastest.

Read and change the code, run commands, provision and query the database, work with your GitHub repository, deploy and roll back, read the logs when something breaks, set DNS and certificates, run the mailboxes. All of it inside your own environment and all of it bounded by the permissions you set.

No. There are no keys to generate, no MCP server to install, no cloud credentials to copy between dashboards and no terminal you are obliged to open. Authorizing your AI here is closer to authorizing an app. The terminal is there if you want it, not because you need it.

Two modes and a board of switches. Read-only lets it inspect the app and recommend; auto-approve lets it edit files and run commands without asking every time. The rows marked enforced are applied outside the model — it is not asked to behave, it is not given the tool. You can change the level of access at any time, cap what one instruction may spend, and read back what was changed and what was run.

Either. Write a prompt and the council pressure-tests the idea, names it, finds the domain and builds it. Or bring the domain and the codebase of an app you already run, connect a Brain and carry on from here.

It keeps running, and so does the AI. Signups, users, checkout, mailboxes, domains, deployments, logs and monitoring live in the same environment as the code, and they are all still there after launch — which is when most of the work actually starts.

Yes, and many people do. Web hosting, mail on your own domain, a database, a runtime, a self-hosted tool, a game or voice server: each is its own group in the shop, priced on its own, and you pick what goes on the machine at checkout. The app plans are the other product — the same machines plus a council that builds and runs an app on them — and you do not need one to get the other.

Yes. Connect a GitHub repository and the container pushes and pulls like any other checkout, so you can work locally while your AI works on the machine against the same history. The browser editor is there for when you do not want to clone anything, not as the only way in.

Your app runs in its own container, not alongside anyone else's. Secrets are encrypted, scoped to what needs them, and revocable. They are never written to logs, and every action taken in your environment is recorded where you can read it back.

You take the domain, the code and the data with you. They were yours throughout, and there is no stage at which we hold them hostage.

A ticket reaches us from any plan in the shop. On the app plans it is written down: Starter is community support, Operator adds email, Studio puts you ahead of the queue. The council is also reading the machine's own logs the whole time, which is usually what answers the question first.