AI Generation
The AI engine generates new components from scratch or modifies existing ones from a natural-language description.
How it works
- Describe what you want — a new component, or a change to the current one
- Pick a model or let Auto routing choose one for the task
- The AI returns the changed files as a diff; the live preview updates instantly
- Accept the changes or revert — nothing is committed until you decide (auto-accept is available in settings)
Models
| Band | Models | Access |
|---|---|---|
| Free model | Gemini 3.5 Flash-Lite | All plans — Free plan: 30 credits / month, max 10 generations / day |
| Budget | Gemini 3.5 Flash-Lite, Gemini 3.8 Flash, GPT-5.6 Luna | Atheros AI Pass and Atheros AI Ultra |
| Premium | Claude Sonnet 5, Gemini 3.1 Pro, GPT-5.6 Terra | Atheros AI Pass and Atheros AI Ultra |
| Frontier | Claude Opus 5, GPT-6 Astra, GPT-5.6 Sol, Fable 5.1 | Atheros AI Ultra |
Besides picking a specific model you can use presets: Auto (classifies the task and balances it across providers, weighting your preferred family from settings — pick “No preference” to route on merit alone), Budget (cheapest capable model), or Ultra (best available model).
Images as input
Attach a screenshot or design image to your prompt — the AI uses it as a visual reference for generation or modification.
Writing good prompts
| Vague | Better |
|---|---|
| "Make it nicer" | "Increase card padding, soften shadows, and round corners to match the rest of the design" |
| "Add a chart" | "Add a bar chart showing monthly revenue below the stats row, using the existing color tokens" |
| "Fix the header" | "Make the header sticky with a blur backdrop and collapse it to 56px on scroll" |
Reliability
If a generation produces a runtime error in the preview, the workspace automatically asks the AI to repair it (up to two attempts). Your chat history is kept as context, so follow-up prompts can build on previous ones.