Goose! Goose! Cook! - Your Voice Sous-Chef

Goose! Goose! Cook! treats a recipe as a graph of steps and a kitchen as a set of limited resources (burners, woks, pots, people). It works out the fastest plan for what you actually have, then runs that plan live, by voice, so nobody has to touch a screen with wet, garlicky hands. Say what you want for dinner and Goose builds the whole session from there: an LLM drafts a step graph per dish, a second model reviews it, and each cook reads a line aloud to bind their voice. The plan itself comes from a branch-and-bound solver, not just a checklist, and it plays out one of two ways. Co-op is one shared timeline, replanned live every time someone says "done." Versus deals out an opening hand and turns the rest of dinner into a real game: steps are up for grabs, cooks claim them by voice, and points are scored by difficulty. Either way it ends in a keepsake summary card with a photo and the night's stats. AssemblyAI's Streaming STT runs the entire pipeline end to end, from the first conversation through the live cook to the summary card, so every stage is navigable hands-free. Mid-cook, "Goose, done with the onion" goes through the LLM Gateway to a tool-calling agent (GPT- 4.1), and the server checks every proposed action against the live run before it applies, so a bad guess never corrupts the plan. Gemini and Claude split the rest of the gateway's work: drafting each dish's step graph, reviewing it, writing the closing summary. A local keyword grammar keeps commands working if the agent itself is slow or down. Goose replies with a local Kokoro TTS voice on WebGPU, with a Web Speech fallback, and an optional speaker-ID sidecar tells two cooks apart on one shared mic. It starts from three enrolled lines and keeps learning from turns it's already sure about, so it gets better as the dinner goes on.