The pipeline is on your Mac, end to end
Speech recognition is whisper.cpp. The voices are Kokoro-82M on the Apple Neural Engine. Summarising and working out which session you addressed is a local Llama 3.2 3B under Ollama.
That is why there is no key to supply, no per-token bill, no rate limit and nothing to leak: there is no server in the loop for anything you say. The app reaches api.novahud.app for licensing, the trial clock and sign-in, and for nothing else.