voice-browser — Jev decides, Playwright acts

model last ms · p50 ms calls 0 · actions 0 cost $0.000000 word→decide ms · →done ms
connecting…

microphone off

Say something like “go to wikipedia”…
Mic needs Chrome/Edge (Web Speech API). Partial transcripts stream to the server as you speak; each update = one Jev request (stale ones are cancelled). Local mode acts after you finish speaking (CPU inference, ~3–5 s) and sees only the top of the page.

action log

jev decision

waiting for speech…
Policy gates (value vs threshold):
gatevaluethresholdnote

intent

yes/no signals

scroll amount

target element

site

text to type / search (verbatim candidates)

controlled page

url
title
site
search box

elements sent to Jev