jev-voice, Browser Extensions and macOS Loops

A recurring shape shows up across this series: strip a task down to "which of these N options is correct?", hand that question to Jev, and let deterministic code do everything else. Three more projects apply the same shape to different surfaces. jev-voice turns spoken commands into macOS actions using local speech-to-text plus one Jev call per command. JevBrowserExt ports the browser-agent pattern from jev-ultrafast into a Manifest V3 Chrome extension. jev-macos-loop drives native macOS automation — Finder, settings — from local OCR and accessibility data, with Jev choosing which control to act on.

All three publish real numbers rather than vague claims of speed, which puts them ahead of much of the field covered so far. None of them have been independently re-run for this article; the figures below are the authors' own, from their own setups.

Key Takeaways

• jev-voice's numbers are unusually granular: ~80–130ms local transcription, ~170–420ms for the Jev call, and an estimated ~1 cent per complex multi-step task.
• JevBrowserExt keeps the exact same one-request-per-step design as jev-ultrafast, just inside a Chrome extension shell instead of a standalone browser session.
• jev-macos-loop scopes its own evidence narrowly: small Finder and settings demos, explicitly not broad reliability evidence for arbitrary desktop tasks.
• All three separate "choose" from "generate." Text typing, when needed, comes from a separate small model or from the user, never from Jev.
• None of these run without local infrastructure — whisper.cpp, an accessibility permission grant, or a Chrome extension install are all prerequisites, not optional.

jev-voice: whisper.cpp locally, one Jev call per command

jev-voice (MIT, 82 stars as of September 2026) lets you talk to a Mac: open apps, type text, search sites, scroll, press keys, control volume and media playback, all by voice. The pipeline runs whisper.cpp locally on Apple Silicon's Metal backend for speech-to-text (reported at roughly 80–130ms), then makes a single Jev API call (reported at roughly 170–420ms) to decide what the transcript means. macOS Accessibility APIs execute the resulting action, and text-to-speech comes from either the macOS say command or ElevenLabs.

The design principle, per the README, is "selection over generation": Jev evaluates several typed multiple-choice questions in parallel per command — which of 13 action types, which installed app, which website's search engine, which candidate text span from the transcript — rather than generating a plan. A default confidence threshold of 0.35 blocks uncertain actions from running. The README's worked examples: a full voice-activation-to-execution loop at roughly 550ms VAD + 100ms transcription + 250ms Jev + 75ms execution; a multi-step AutoTrader browser task taking about 13 seconds and around 30 Jev calls; and chess moves proposed by Stockfish and filtered through Jev selection at about 1.3 seconds per move. Estimated cost for a complex task is about 1 cent. Multi-step desktop runs cap at 40 actions, and three consecutive actions that don't change screen state trigger failure detection. Setup requires macOS on Apple Silicon, a uv-managed Python environment, a TypeSafe API key, and permissions for Microphone, Accessibility and Input Monitoring — the installer also remaps Caps Lock to F18 as an activation key. Aiskill.market entry: /skills/jev-voice-kevinbadi.

JevBrowserExt: jev-ultrafast as a Chrome extension

JevBrowserExt (MIT, 21 stars as of September 2026) is described in the project's own listing as a Manifest V3 Chrome port of jev-ultrafast — same one-request-per-step design, new delivery shell. Jev picks the operation (CLICK, TYPE_TEXT, SELECT, SCROLL_DOWN, SCROLL_UP, PRESS_ENTER, WAIT, DONE, or BLOCKED) and the matching DOM element in a single request per step; PRESS_ENTER runs as a separate key control, and separate yes/no checks cover whether the goal looks complete and whether the run looks stuck. A small chat model fills in TYPE_TEXT only when text actually needs typing, mirroring the read/pick/execute split used across this series.

Running inside the user's own browser tabs means no screenshots are captured — the action, its target, and any typed string stay inspectable in the extension's own logs. The project's listing is careful to note what its own checks establish: source-level Jev integration, not runtime, safety, or performance validation. That's a meaningful caveat — a code review confirming the wiring is correct is different from confirming the extension performs reliably across real sites. Aiskill.market entry: /skills/jevbrowserext-chy4pro.

jev-macos-loop: local OCR for Finder and settings

jev-macos-loop (AGPL-3.0, 20 stars as of September 2026) is a native macOS automation loop that reads the screen with local OCR and accessibility information, then has Jev select which control or group to act on from that set of locally-observed candidates. Execution — clicks, coordinates, actual input — stays entirely on the Mac; Jev's job is narrowed to picking from what's already been extracted, the same pattern as typesafe-computer-use but scoped to Finder and system settings rather than the whole desktop.

The project's own evidence framing is unusually specific about its limits: it offers small Finder and settings demos, and explicitly states these are not broad reliability evidence for arbitrary desktop tasks. That's a fair line to draw at 20 stars — there's real, inspectable candidate-extraction and execution-check logic here, but no claim that it generalizes beyond the demonstrated workflows. Aiskill.market entry: /skills/jev-macos-loop-jcpsimmons.

What to do next

1. If you want voice control today, jev-voice has the most complete numbers in this trio — but budget setup time for whisper.cpp, permission grants, and the Caps Lock remap before expecting it to work out of the box.
2. Treat JevBrowserExt's "source-level validation" language precisely. A correct integration isn't the same as a tested one; run your own tasks before relying on it.
3. Keep jev-macos-loop to Finder and settings-style tasks until you've verified it on your own workflow — the author doesn't claim more than that.
4. Watch the confidence thresholds in each project (0.35 for jev-voice, separate goal/stuck checks for JevBrowserExt) — they're the actual safety mechanism, more than any marketing claim.
5. Compare designs across the desktop cluster. For the OCR-plus-accessibility-tree approach at full desktop scope, see typesafe-computer-use and Jev-cu; for evaluation method, see how to evaluate Jev projects honestly. The Awesome Jev radar tracks new entries as they appear.