Chervil already lives in your tray, a global hotkey away. Now it can also just listen — say the wake word and the quick-ask bar appears, ready to take a spoken request. No hotkey, no clicking, no reaching for the mouse.
On-device by default
The part we care most about: wake-word detection runs entirely on your machine. Your microphone audio never leaves the device until you actually speak a command — and even then it only goes to the voice service you configured. There's no always-on stream to anyone's cloud.
Under the hood it's openWakeWord running on onnxruntime-web — free, open-source, and crucially no account and no API key. (Our first cut used a commercial engine whose free tier is being retired; moving to openWakeWord keeps the feature free and private, with nothing to sign up for.)
Turning it on
In Settings → Listening, flip on listening mode and pick a wake word. Three work out of the box — Hey Jarvis, Alexa, and Hey Mycroft — so you can try it in ten seconds.
Making it "Hey Sprig"
Want the real thing? openWakeWord can train a custom word for free. Open the openWakeWord training notebook, set the phrase to hey sprig, run it on the free GPU (about half an hour, no recording needed — it synthesizes the samples), and download the resulting .onnx. Then load it under Settings → Listening → Custom model. We're bundling a "Hey Sprig" model as the default in an upcoming build so it works the moment you install.
It's a small feature with a clear rule behind it: the convenience of talking to your browser shouldn't cost you your privacy.