Question 1

Can it type arbitrary search queries and URLs, or only fixed ones?

Accepted Answer

Jev never writes text. Search queries, typed text, and URLs are extracted as candidate spans by regex in code, and Jev only picks one, which is copied verbatim. Navigation uses a site list plus domain-looking spans from what you said, so nothing is invented.

Question 2

Will it mis-hear me and click something destructive?

Accepted Answer

Policy requires a destructive probability of 0.5 or more on a click before it asks you to say confirm or cancel. The README still calls that a convenience rather than a guarantee, and warns against logging into accounts in the persistent .browser-profile/ directory that the controlled window uses.

Question 3

Does this work in Firefox or Safari at all?

Accepted Answer

No, because the microphone path uses the browser Web Speech API, which is not available in Firefox or Safari. You need Chrome or Edge for the control page. The controlled browser is a separate Chromium window Playwright launches, though you can also attach to a Chrome you already started with a debugging port.

Question 4

Will it start acting before I finish my sentence?

Accepted Answer

Yes, that is the design. Every partial transcript, debounced 200 ms, sends one request with nine to eleven typed questions including a completeness question, and policy acts once completeness clears 0.6, after 900 ms of silence, or on the recognizer's final result. Free-text intents wait longer so a query is not truncated mid-phrase.

Question 5

How do I run voice-browser?

Accepted Answer

Requires Node 20 or newer and Chrome or Edge for the microphone. Clone the repo, `npm install`, `npx playwright install chromium`, copy .env.example to .env and paste a TypeSafe key, then `./run.sh`. Open http://localhost:8787 in your normal Chrome and click Start mic.

Question 6

Where does the API key live, and who can reach it?

Accepted Answer

The key is only read by the Node process, and the control page never sees it. The server listens on 127.0.0.1 only; anyone who can reach the control port can drive the browser and spend your credits, so only use --host 0.0.0.0 on a trusted network.

Question 7

How does it remember earlier commands?

Accepted Answer

Jev has no memory between requests, so the state carries the page you came from and the last three executed actions: what you said, what was done and what happened. That is what makes 'go back to the results', 'not that one' and 'open its documentation' resolvable.

Voice Browser

About this project

Highlights

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community