No backend, no API key, no data leaving your machine. A small language model (Qwen2.5 0.5B, or a compatible 1B model on GPUs without f16) loads straight into your browser over WebGPU and runs on your own hardware. First wake downloads the weights (≈0.4–1 GB, cached after). Then ask the night editor something — about the stack, the city, or the meaning of a deadline.