This website will load a choosen language model, store it in browser cache, and then generate text based on the provided grammar. It uses Low-level Guidance (llguidance) library compiled to WebAssembly and a bunch of glue code. The llguidance grammars are generated from TypeScript/JavaScript code via guidance-ts library.
To simulate proposed APIs that are not yet available in the browser, the site uses WebLLM (see implementation). WebLLM requires a GPU exposed via WebGPU. It seems to work in Edge on Mac and Windows.