Agents: read the raw Markdown of this page, or start at llms.txt.
Builds: desktop, mobile and voice computer use
TL;DR 10 builds plus names-only lines (split from Builds: browser, computer use and interface on 2026-09-25). Local code reads the screen (accessibility tree, OCR, on-device detection), Jev picks the next control or command from that list, and code acts and checks the result. The best evidence is small: Jev macOS Loop's 6/6 native tasks, with its failures written down. Policy lives in code: macbrow's safety file exists because an early build emptied a Desktop into one folder. Several builds here read other people's messages; see the flags.
How to read. Builder reports = the builder's own numbers, one run on their own machine, unverified. Pattern IDs: P12-P14 in Patterns: browser, computer use, voice and product UI; P02-P06 in Patterns: agent internals, routing, gates, context and memory; P38 in Patterns: gates, simulation, personas and other shapes (P38+). Browser agents and UI elements: Builds: browser, computer use and interface. Point these agents only at apps and accounts you may automate (AUP note on Patterns: browser, computer use, voice and product UI).
Desktop and mobile control
| Build | Jev decides | Builder reports | P |
|---|---|---|---|
| Jev macOS Loop (Josh C. Simmons) | Local perception (ScreenCaptureKit, OmniParser CoreML, Apple Vision OCR, accessibility labels) turns one window into element ids; Jev picks the next click from that finite list, text only (pixels and coordinates stay on the Mac); code checks focus, occlusion and a fresh frame before input; a separate verifier, not Jev's DONE, decides success. Vercel Gateway (default), OpenRouter or direct | Measured 2026-09-17/18: 6/6 native GUI tasks on Vercel and 6/6 on OpenRouter; median decision 301.9 / 293.5 ms (p95 547.2 ms over 22 Vercel decisions); Finder: 9 files sorted in 7.39 s with one classification request and three group drags, vs 22.69 s one by one. Documented failures: a later Calculator regression typed an extra digit (3,968 instead of 448), caught by the independent verifier; OCR disagreed with accessibility labels or lagged a transition; an early suite scored 5/6 (action label vs checkbox state). Author: small samples, not reliability evidence; the direct TypeSafe route has no live test. No typing, one window. AGPL-3.0 | P14, P04 |
| macbrow (Pratim) | Voice (Gradium speech in and out, via LiveKit) → one Choice over the tools for the apps open now plus "chat" and "new action"; each candidate's enum arguments asked speculatively in the same request, free text as a span of the utterance; web tasks go to jev-ultrafast in your Chrome; follow-up, completeness and "did the script do it" are single Noul/Choice checks |
~300 ms routing; Amazon add-to-cart 4 steps in 11 s; a follow-up 6 steps in 3.7 s; a trailer playing 1.6 s after speech ends (author, own machine). Failure that shaped it: an early build asked to clean up the Desktop moved every Desktop file into a folder; policy.py now blocks file moves and deletes, settings, terminals, purchases and sign-ins, and state changes need a spoken yes (a misheard valid command is not covered). Flag: for unknown actions an LLM writes new AppleScript tools (compile, effect, policy and Jev-review gates), cached and run locally. Author: experimental, not for production. MIT |
P14, P03, P02 |
| Jev-cu (Sac-Y; the code behind @Saccc_c's "Jev Use" post) | Codex Computer Use reads the accessibility tree; Jev picks element, action, done-ness and risk; dry-run by default; delete, send, pay, install and similar stop at confirm; per-app allowlist in scripts/policy.mjs |
Post (video): smoother than Codex's own computer use at similar token use, unverified: an offline eval (npm run p0) ships without committed results. Goals written in English because Jev is most accurate there (matches Models, aliases, pricing, rate limits, context: verified). MIT |
P14, P03 |
| jev-use (savka777; not shitianfang/jev-use on Head-to-head: Jev against other models and methods) | macOS voice or typed commands over the accessibility tree: one request per step picks operation and target; typed text is a span of your sentence; low-confidence and destructive picks stop and ask; optional "Hey Jev" wake phrase | 0.3-1.5 s per step (author). The README lists exactly what is sent: no screenshots, secure fields excluded. MIT | P14, P38 |
| agent-device QA (Mike Grabowski, Callstack) | Mobile accessibility snapshot → choice over press/fill/scroll/wait plus pass/fail/incomplete; text to type is baked into the option |
14 s run, $0.0023. systemOne + choice() matches JavaScript/TypeScript SDK quick contract: install, client, choice/score/noul, answers, errors (the page a builder reads) (verified) |
P14 |
| Screenshot-free computer use (@milindlabs) | On-device segmentation + OCR turn the screen into a text list; Jev picks what to click, in a loop | ~90 ms per decision | P14 |
| WeChat group bot (reviewed by @VersunPan) | Jev + DeepSeek over the Android accessibility tree, screenshot + local OCR fallback, last 10 messages | No numbers; warns WeChat can detect it and ban the account | P14 |
| Yappy (Mitosis Labs; closed, paid) | Mac computer-use agent: the front window as an accessibility table of controls (no screenshots) → one operation question and a target question per operation; a chat model types; hands off to the full agent below 0.25 confidence, after three no-op actions, or when a field value is unknown | Vendor's own single run (2026-09-18, one job-application form, n=1): 1m54s vs 13m14s and ~$0.24 vs ~$4.24 for a named competitor, whose time includes ~2m20s waiting on the user (research page); five Jev decisions, 275-690 ms each, $0.003 in total | P14, P06 |
Voice commands
| Build | Jev decides | Builder reports | P |
|---|---|---|---|
| Jev Voice (kevinbadi) | Local whisper.cpp transcribes everything; only utterances naming the assistant reach Jev (a wake gate); one fan-out request picks the action and its arguments from code-made candidates; multi-step tasks run the jev-ultrafast loop on the accessibility tree or in Chrome | Jev fan-out 170-420 ms, whisper 80-130 ms (author's table); an AutoTrader search ~13 s, ~30 Jev calls, about one cent. Setup remaps Caps Lock with a LaunchAgent. Flag: the repo also drives rated chess.com games with Stockfish proposing and Jev picking; its README reports wins in rated games and itself notes engine play in rated games breaks chess.com's fair-play rules. Not a use to describe or recommend. MIT | P14, P38 |
| Live Jev (Okina Audio) | One English or Japanese sentence → an allow-listed Ableton Live action, candidates taken from your own set and plug-ins; fixed phrases handled locally without Jev; an optional Gemini rewrite is checked against the set and always confirmed | Chains of up to four commands are all checked before any runs, and earlier parts are rolled back if a later one fails; undo only with a receipt; a named track that does not exist writes nothing. Source only. MIT | P14 |
(jev-canvas, voice plus a pointing fingertip on a canvas, is on Builds: browser, computer use and interface with the interface elements.)
Also seen, names only: jev-cua (ronadin2002; floating macOS bar, accessibility controls → Jev choice → act → re-observe; demo, no licence), jev-desktop (yikangy873-gif; bounded Jev loop inside Codex Computer Use, every executable control allowlisted by the calling task; MIT). Messaging readers, flagged: JevBystander (Android accessibility service reading private WeChat chats; ships a debug-signed APK with a public keystore: do not install), its sibling JevIntent (Nisaka520; long-press a WeChat message for intent, emotion and urgency; hooks WeChat through LSPosed and sends the other person's messages out) and Yanwai (YIRC99; emotion and subtext under WeChat bubbles; sideloaded APK hooking WeChat, no licence, third-party messages sent out).
What these builds teach when advising
- Verify outside the model. Jev macOS Loop's verifier, not Jev's DONE, caught 3,968 ≠ 448; Jev Voice and jev-use also call DONE a claim, not proof.
- Write the policy before the first run. macbrow's Desktop incident; Jev-cu defaults to dry-run; Live Jev runs a chain all-or-nothing with rollback.
- Gate what reaches Jev. Wake words (Jev Voice, jev-use) and local fixed phrases (Live Jev) keep stray speech from becoming actions.
- Vendor numbers are n=1. Yappy's comparison is one run on one form, by the seller.
- Other people's messages are not yours to send. The WeChat readers send third-party chats to an API without that person's consent, and client hooking can get an account banned.
Related
- Patterns: browser, computer use, voice and product UI — P14 with its fit verdict and the AUP note
- Builds: browser, computer use and interface — browser agents and in-UI classifiers
- Builds: coding agents, harnesses and orchestration, Builds: games, simulators, robots and devices — coding agents; games, robots and devices
- Failure reports: where Jev broke, lost, or was the wrong tool — where Jev broke
Sources
Links are inline in each row; the raw captures are listed in the frontmatter. Names-only lines rest on our vetting at source where no capture exists: 2026-09-24 (docs/sweep/2026-09-24-awesome-jev/part1.md, part2.md) and 2026-09-25 (docs/sweep/2026-09-25-mrjev/partb.md).