Local AI — check settings
Paste a Hugging Face model URL or repo id. MLC/WebLLM packages and Transformers.js ONNX models are supported; the repo is inspected and incompatible repos are rejected with an explanation. Downloads come from Hugging Face — inference still runs locally on your GPU.
VRAM figures are estimates derived from model size and quantization, not measured usage. Token speed is measured live during generation.
Run node local-bridge.mjs to let the AI read/write files in bridge-workspace/, run shell commands (each command asks for your approval), and serve your app on your Wi-Fi for phone testing. Also paste any MCP server HTTP endpoints below — their tools are merged into the AI's toolset.
Publishing uploads your project files (not your settings or keys) to your Supabase app store. Anyone can install from the store.
A snapshot is taken before every AI edit and after periods of manual editing. Restoring rolls the whole project back.
Edit freely. Starting the build runs Architect (plan) → Builder (code) → Tester (verify) using your local models.
This uses the local bridge: your project is served over your Wi-Fi so any phone on the same network can open it. If the QR area is empty, run node local-bridge.mjs next to the IDE and paste the token in Settings.