Build with your AI partner

Describe a concrete outcome, include the framework and version you target, and explain how you expect the project to behave.

Four ways to work

• Build: propose complete project files or a new feature.
• Debug: investigate supplied errors and propose repairs.
• Explain: explain architecture or code without proposing files.
• Optimize: improve maintainability or performance.

Review every change

The agent returns an explanation and proposed files. Select a file to inspect its diff. Accepting applies the proposed files to the project; rejecting leaves the project unchanged.

If the project changed while the request was running, acceptance is blocked to avoid overwriting your edits. Save your changes and request a fresh proposal.

Credits and reservations

The platform reserves a maximum number of credits before placing the request on its queue. Your balance cannot fall below zero. The final charge uses provider-reported input and output token usage and the model's configured rates. Unused reserved credits are returned.

A failed request returns the reservation. Cancelling a queued request returns it immediately. Cancelling a running request discards the proposed changes when the provider finishes; actual usage is still charged. Running cancellation does not terminate the provider's HTTP request.

Current limits

Projects support up to 100 files and 500 KB of saved source. AI context is limited to 60 KB, including recent conversation history. The workspace polls for completed responses; live token streaming is not yet available.

Generated code is untrusted and is never executed on the application server. No test execution or sandbox preview is currently available. Validate and test exported projects in a separate isolated environment.