User repository on GitHub
One TypeSafe Jev decision per step over the A11Y tree, executed via ADB. No screenshots and ultra fast!
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
Jev Mobile — xinwang-nwpu
An Android automation agent that decides from the accessibility tree instead of screenshots: one TypeSafe Jev request per step picks the operation and target, ADB executes, and an optional vision fallback takes over when the fast path gets stuck.
- Stars
- 8
- Language
- Python
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- xinwang-nwpu
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Jev Mobile — Friedjof.
Highlights
- Each observation produces an indexed A11Y element table; one Jev request chooses the operation, click target and type target together.
- The demo task — playing a specific video on Bilibili — took 18 seconds total, with 2–3 seconds per action decision and no screenshots.
- Completion is a separate Noul question in the same request; a DONE veto triggers WAIT and re-judging before the vision chain is used.
- Optional vision fallback (OpenAI-compatible, e.g. OpenRouter) shares 60 action attempts and 240 model call attempts with the fast path.
- Vision-only mode runs from the first step without a Jev or text-assistant key, using the --vision-only CLI flag or a config setting.
Quickstart
bash
python -m jev_mobile --vision-only --task "你的任务"Watch out
Copy reviewed against the truncated repo README on 2026-10-15; the capture stops inside the 'why it's fast' section, so setup details after that were not reviewed. The 18-second demo and 2–3s decision timings are the author's own single-task measurements, and the README is written in Chinese with an English version linked.
Questions about this entry
Short answers drawn from the linked project or page.
An Android automation agent: you give a natural-language goal, Jev picks the operation and target from the accessibility tree, and ADB executes taps, text input, scrolls and navigation — without screenshots in the default loop.
One request per observation answers operation, click target and type target together — two decisions, one network round trip. The completion question, whether the goal is achieved, is a separate Noul evaluated in the same request.
The README's demo — playing a specific video on Bilibili — finished in 18 seconds total, with each action decision taking 2–3 seconds. It reports no other benchmark, so treat those as single-task measurements.
An Android device connected over ADB with USB debugging, Python, and a TypeSafe key for the Jev fast path. The README notes a Jev Bridge (ai.jev.bridge 0.1.4+) reads the whole tree and phone state via ADB forwarding.
When the fast path gets stuck — A11Y read failure, no target, model failure, three unchanged steps or repeated button bouncing — an optional OpenAI-compatible vision model takes over, sharing 60 action attempts and 240 model call attempts with the fast path.
The fast path asks a Noul question and accepts ≥0.5 as done. Vetoed DONEs go through WAIT and re-judging; after two vetoes the third DONE goes to the vision chain, where a separate Verifier checks the original task checklist.
It is a young Python project, created 2026-09-21, whose README is in Chinese, with no published benchmarks beyond the demo. The captured README was truncated before the end of its performance section, so later details were not reviewed.
More Jev resources like this
A dependency-free Python pipeline that rebuilds a site's internal link map with Jev: code picks candidate targets and in-copy anchor phrases, Jev judges them, and only links that clear the threshold are placed.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
stas4000/jev-linkmapHTML
stas4000/jev-linkmapHTMLTraffic signal control with Jev applied to SUMO simulation: each control step packs the road network into one state with a choice question per intersection, and the model's chosen phase, distribution and confidence drive the signal with local-ranking fallbacks.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
usail-hkust/jevlightPython
usail-hkust/jevlightPythonA C/C++ compiler driver that combines Clang's optimiser with Jev for function-call inlining at every discretionary call site. The project reports 58.2% code-size reduction on Statemate from the Embench benchmark; Linux with pinned LLVM 21.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
ramneet-singh/jevoptPython
ramneet-singh/jevoptPythonA source-visible Android vision control prototype where a vision agent reads screenshots, a text-only supervisor plans, and TypeSafe JEV chooses among candidate actions; an ADB executor runs them under a local web console.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
ztrrtuo/jev-phonecontrolPython
ztrrtuo/jev-phonecontrolPythonMaykana's free tool that lets a cheap AI test a web app like a person: write checks in plain sentences, Jev picks the next step from a short list, and below 80% confidence it stops and hands over to a human instead of clicking. A DeepSeek vision pass reports visual defects.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
elshinq/jevaluateJavaScript
elshinq/jevaluateJavaScriptLocal, inspectable image perception for the text-only Jev: OCR text in reading order plus the same words on a character grid become the state Jev decides over, returned as a plain dict. CLI, MCP server and agent skill.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
leddoengano/jev-eyesPython
leddoengano/jev-eyesPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.