Most AI app builders write files and tell you afterwards. You send a prompt, the agent works, and some time later you have a diff to audit — or worse, you don't, and you find out what changed when something breaks.
We built ours the other way round: the agent proposes a file, and then it stops.
Here is a full run, one sentence in to a working landing page out:
What the approval gate actually does
Every write_file and edit_file is intercepted before it touches disk. You get the path, the diff, and three choices: accept it, reject it, or approve the rest of the run so it stops asking.
The escape hatch matters as much as the gate. A review step you cannot turn off is a review step people click through without reading — so "Approve the rest" is a first-class button, not something you find by giving up.
The part I did not script
Around 1:12 in the video, the agent's own build check fails. It reads vite.config.ts, notices the plugins array is [react()] with no Tailwind plugin, runs a dry-run edit to confirm the string it wants to change exists, applies the fix, and then notes that Vite config changes are not picked up by HMR — so it restarts the dev server and verifies the CSS is actually being served.
That is the real trajectory, not a demo edit. It is also the strongest argument for the gate: you can watch the reasoning that produced each change.
Why we charge for the sandbox and not the tokens
The agent runs in a real Kubernetes workspace you can open in VS Code. Bring your own API key — or point it at a local Ollama — and the tokens cost you nothing, because they never go through our provider account.
What you pay for is the machine it runs on. No credits, no overage, and a failed build does not spend one of your runs.
What comes out is an ordinary Vite or Next project with a normal package.json. No proprietary format, nothing that needs us to keep running.
Try it
Free tier, no credit card: https://nocoder.codes/?utm_source=devto&utm_campaign=demo
If you try it and the agent does something dumb, I would genuinely like to see the trajectory — that is the fastest way we fix things.
Top comments (0)