Murmur is a complete voice-assistant front end built on the browser's own Web Speech API - no keys, no servers, no SDK. Hold the mic button (or Space) to talk, tap to toggle, or switch on hands-free mode with an optional "Hey Murmur" wake word. Replies are spoken aloud with speechSynthesis, with a voice picker plus rate, pitch and volume controls.

The centrepiece is an animated canvas orb with a radial waveform. It idles, listens (optionally reacting to your real microphone level), thinks and speaks, and respects reduced-motion preferences. Around it sit a live transcript, notes, running countdown timers with a chime, and ready-made intents: timers, notes, a deterministic weather mock, jokes, time/date and spoken maths.

The brain is isolated in js/ai.js behind one function, MurmurAI.respond(). The default is a rule-based local intent engine; a documented OpenAI-compatible example shows how to hand commands to a real LLM using a key that lives only in the visitor's localStorage. If speech recognition is unavailable the app degrades gracefully to typed commands, and says so.