OpenStitch is a locally-runnable AI design tool that generates rendered, interactive frontend code from text descriptions, screenshots, or wireframe sketches. It runs entirely on your machine via Docker and Ollama - no data leaves the box. The canvas lets you arrange multiple screens, wire them together with navigation flows, and test click-through behavior in Play Mode.
AI-assisted UI generation tools are almost all cloud-only and subscription-gated. They also tend to produce isolated components rather than coherent multi-screen prototypes. There is no good open-source equivalent that runs locally, supports vision models for screenshot replication, and understands design system context across screens.
Each screen in the canvas shares a DESIGN.md - a short spec that anchors visual consistency across generated screens without manual stitching. Vision models (Qwen2.5-VL via Ollama) can replicate a screenshot or interpret a sketch; text models handle natural language specs. The model tier is configurable: default to a local Ollama instance, fall back to OpenRouter for cloud access.
50+ pre-built design system templates (Linear, Stripe, Vercel, and others) mean the output starts from a coherent visual language rather than generic Tailwind defaults.
React 18 frontend on a React Flow infinite canvas with CodeMirror for live code editing and WebSocket-powered preview. FastAPI backend with async SQLite, structlog, and an Nginx reverse proxy so everything runs behind a single port. Docker Compose handles the full stack. MIT licensed.