What it is
Voice-driven browser use in your signed-in profile.
What it does: A native host plans with Claude, then Jev repeatedly chooses click/type/select/scroll on a fresh element snapshot until the goal is done or blocked.
Why it belongs in the showcase: Combines voice intent, stepwise browser action selection, and an irreversible-action gate without attaching a debugger.
Caveats: No published success-rate table; site compatibility limits apply.
Evidence
Public MIT repo. README documents a four-step loop (Deepgram live transcript → claude -p plan → Jev action loop via browser-use/jev-ultrafast style element lists → optional Claude answer), measured end-to-end examples in about 9–19 s, irreversible-looking clicks gated by Approve/Decline, keys in ~/.config/jev-voice/env mode 600, and no remote-debugging port.
Caveats
- Synthetic events have isTrusted=false; some widgets ignore them
- Depends on Claude CLI login plus a TypeSafe key; no independent task-success benchmark in the README
- Approval verb lists cover main EN/FR/DE/ES forms but are not exhaustive
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.