jev-voice-browser
Talk to a real Chromium window; Jev scores intent and targets in ~300 ms per word and Playwright acts—often before you finish the sentence.
Moritz Kremb open-sourced a Node app that controls a headed Chromium window by voice.
Speech streams word-by-word from Chrome's Web Speech API to a small Node server. On every partial transcript the server asks Jev (jev-1.13.0) typed questions—intent, target element, site, whether the command is complete, whether it's destructive—and gets probabilities back in ~250–350 ms. Code decides whether to act, wait, ask, or ignore.
Jev never generates text. Search queries, typed text, and URLs are extracted as candidate spans by code; Jev only picks one, copied verbatim.
Run it
git clone https://github.com/moritzkremb/jev-voice-browser.git
cd jev-voice-browser
npm install
npx playwright install chromium
cp .env.example .env # TypeSafe API key
./run.sh # http://localhost:8787
Open the control page in Chrome, start the mic, and speak. A separate Chromium window acts on your commands. Real API calls cost about $0.0002 each.
What you can say
- "go to wikipedia" / "open youtube"
- "search for alan turing"
- "click the first result"
- "scroll down a bit" / "go back" / "reload"
- "click place order" → asks you to say confirm