Project · local-first AI spatial workspace
A canvas where every document is still a real file.
Documents, spreadsheets, code, PDFs, calendars and 3D models live on one infinite canvas as
fully editable items, in two-way sync with files on disk. A Rust/WebAssembly engine draws the canvas on the
GPU, and the AI (semantic search, OCR, speech-to-text, a local LLM) runs on the machine, offline.
- GPU renderer. wgpu/WebGPU with a WebGL2 fallback, drawing 1,000+ rich items at a 7.1 ms
typical frame, with pan and zoom multitudes faster than the JavaScript DOM renderer it replaced.
- On-device AI. Semantic search on ONNX Runtime Web, OCR, Whisper speech-to-text and a llama.cpp
assistant, all in Web Workers so the canvas never drops a frame.
- Files stay files. Dropped files stay linked to their items; items made in the app get real files
on disk; dragging one out opens it in the right editor.
- Built to stay honest. A pnpm monorepo with 5,000+ automated tests, CI and performance
regression gates. Yjs CRDTs for conflict-free sync.
TypeScriptReactElectron
RustWebAssemblyWebGPU
Yjs CRDTsONNX RuntimeVitest
Theme
Drag the canvas or an itemScroll or pinch to zoom
Click an item to open itSearch to filter
A browser-light recreation, built for this page in about 9 KB of JavaScript. It lays the
cards out at whatever zoom you leave them at, so the type stays sharp, and drops to blocks when
they are too small to read. The real workspace does this through a Rust/WebGPU engine with a
WebGL2 fallback.