Check before you download
The AI tab first runs a capability check: whether WebGPU exists, whether a GPU adapter is granted, whether it supports 16-bit float shaders, its buffer limits, and how much browser storage is free. Each model in the catalogue is marked usable or not on this device, with the reason.
Nothing downloads until you choose a model and confirm. Sizes are shown up front — from roughly 300 MB to 900 MB of weights — along with GPU memory needs, the licence and the source repository. Downloads show progress and can be cancelled; downloaded parts stay cached, so resuming continues where it stopped.
What the model is allowed to do
Plan refinement: the model returns JSON constrained to the plan schema. It is parsed, repaired if the damage is mechanical (code fences, trailing commas, truncation), validated field by field, and otherwise rolled back to the template plan.
Code edits: you choose a file and describe a change. The model’s rewrite is checked for syntax, unsafe patterns and secrets, then shown as a diff. It applies only when you approve it, and Undo reverses it.
Frequently asked questions
Does any prompt leave my device?
No. The runtime and weights are downloaded from public CDNs; those requests contain no prompts or project content. Inference runs locally on your GPU.
Which browsers can run it?
Browsers with WebGPU, such as recent Chrome and Edge on desktop, and Safari versions with WebGPU enabled. Others get the template path.
How do I remove a downloaded model?
Use “Delete cached model” in the AI tab, or clear this site’s data in your browser settings.