The AI agent
It builds the change. Then it keeps doing the boring half forever.
Plenty of things will write you an app. The part nobody wants is what happens afterwards — the upstream release you did not notice, the dependency bump, the merge you keep meaning to do. That is the work the agent is really for, and it is the reason personalizing an app here is not the same as forking one on GitHub.
Implementing
You describe the outcome. It decides what the platform has to do.
- It picks the pieces
- You do not have to know whether your idea needs a queue, a search index or just a table. The agent decides what the app needs and wires it together, so the vocabulary of the platform never becomes your problem.
- It works in your repository
- Changes are commits on a branch of a repository you own, with a plain-language description of what they touch. Review them like any other pull request, or don't — they are there either way.
- It gets a real environment
- The agent works against its own branch of your Postgres, copy-on-write, with credentials scoped to that branch alone. It can run the thing it just wrote instead of guessing whether it works.
Maintaining
Upkeep is a loop, not a favour it does once.
A fork starts rotting the moment you make it. SproutOS merges upstream into your copy on a schedule and tells you which of three things happened — nothing to do, it merged, or you and upstream disagree and here is a pull request.
- On a cadence you pick
- Every upstream release tag, or daily, weekly, or monthly. A missed interval stays due, so downtime catches up rather than silently skipping a week.
- A conflict is not a failure
- When upstream and your change touch the same lines, you get a pull request to resolve. That is the normal state of a fork somebody is using, and it does not count against anything.
- Five real failures in a row pauses it
- Every run costs money, so a fork whose upstream has diverged past reconciliation stops rather than failing identically every night forever. You are told, and you can restart it.
The same loop, drawn end to end, is on the personalization page.
Bring your own model
It can run on tokens you already pay for.
The Claude Code subscription you have, an API key of your own, or a model hosted inside your company. Upkeep is not a reason to buy a second AI budget, and a team that is not allowed to send code to a third party can point this at their own host.
Or ours, at cost
Platform tokens are passed through.
If you use the model we provide, the tokens are billed at what they cost us — there is no platform fee on top of AI usage, and the agent's own running time is not charged at all. What you pay for is the work it causes, not the fact that it thought about it.
How billing works