User repository on GitHub
Fast GUI Grounding with jev
- Forks
- 2
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Tools & CLIs
GUI JEV Harness
Python harness for GUI grounding: it divides a screenshot into grid tiles, has a vision model describe each tile neutrally, then asks typesafe-ai/jev to choose among the descriptions. Local confidence and margin gates decide whether to descend or refuse, and it emits only a point and bounding box.
- Stars
- 4
- Language
- Python
- License
- MIT
- Published by
- Community
- Last verified
- Oct 2026
About this project
An auditable, screenshot-in/coordinate-out harness that uses TypeSafe JEV as a closed-set decision layer during recursive GUI image search.
Quoted from the project's README. Read the full README.
- Languages
- Python
- Platforms
- CLIAPIVercel AI Gateway
- Capabilities
- Open sourceRequires API key
Highlights
- Pixels are never sent to Jev: a separate vision model describes tiles, then Jev chooses among descriptions
- Local gates validate the closed candidate set, numeric types, probability sum, model and argmax; failures fail closed
- Emits only a raster point and bbox with status located, refused, limit_reached or error; it never clicks
- Pinned 12-case ScreenSpot smoke set: 9/12 point hits (text 6/6, icon 3/6) with local OpenJev 0.8B
- Offline replay uses vision manifests and decision scripts keyed by crop; deterministic, no network calls
Watch out
MIT license. Online mode needs a Vercel AI Gateway key and a vision model you choose; the final box is the remaining search region, not a guaranteed control bounding box. OpenJev is an independent classifier, not TypeSafe Jev.
Questions about this entry
Short answers drawn from the linked project or page.
TypeSafe's Jev API defines text and JSON inputs with typed probability outputs, not image input. The harness therefore uses a vision-capable model to describe grid tiles neutrally, and Jev only chooses among those text descriptions.
No. It never captures the live screen or executes clicks. A successful result is a raster point and bounding box, and the caller owns screenshot capture, display scaling, freshness checks and click execution.
At each layer it divides the current crop into a deterministic grid with thin drawn boundaries, zero-based row-major tile ids, and sends the annotated crop to the vision model through Vercel AI Gateway. The vision prompt is goal-blind, so it describes visible text and controls but cannot choose a tile.
Local code validates the closed candidate set, raw numeric types, approximate probability sum, response model and argmax, then applies confidence and margin gates. Failures fail closed, and the final status is one of located, refused, limit_reached or error.
Python 3.11 or newer, a virtualenv, then `python -m pip install -e '.[dev]'`. Online mode needs an AI Gateway key exported as AI_GATEWAY_API_KEY and a vision model id in JEV_GUI_VISION_MODEL, which is deliberately not hardcoded because Gateway availability changes. Locate a goal with rows, columns, max-depth and target size, writing a trace and an annotated crop.
On a pinned 12-case ScreenSpot set of three Windows text, three Windows icon, three iOS text and three iOS icon targets, the 32-pixel configuration reached 9 of 12 point hits with the local OpenJev 0.8B checkpoint: text 6 of 6 and icon 3 of 6. The README calls that a smoke result, not a benchmark score.
The final box is the remaining search region, not a guaranteed control bounding box. Uniform grids can split a target across tiles, duplicate-looking controls stay ambiguous, and a vision description error propagates into the Jev decision, so production callers should re-capture the screen and snap to a DOM, accessibility, OCR or detector box where possible.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsX: Jev driving real Chrome from Grok Bot
Hacker News: Cua asks which computer-use steps need a frontier model and which are fast decisions.
frabonacci95 pointsShow HN: CUA-S1 – A System One Model for Computer UseRead the thread on Hacker NewsX: Voice-controlled computer use on a Mac
More Jev resources like this
3GitHub stars
Go CLI for Jev with noul, choice, score, eval, rank, and models commands: rank sends one Score question per item per weighted YAML dimension in a single request and sums weight times score in code, while threshold flags turn a failed gate into exit code 10 for scripts and CI.
Jev Tools & CLIs#community#cli#go
Community
saembit/jeff-cliGo
saembit/jeff-cliGo2GitHub stars
Zero-dependency Python CLI that routes small semantic judgments to Jev for agents and CLI pipelines: choose (Choice), check (Noul), score (ordered ratings), batch (1-32 questions in one request) and purify (verbatim passage extraction). GPL-3.0 licensed.
Jev Tools & CLIs#community#cli#decision-routing
Community
logicrw/ask-jevPython
logicrw/ask-jevPython110.5kGitHub stars
Multi-agent trading research with optional Jev screening of StockTwits and Reddit posts for company relevance before sentiment analysis runs.
Jev Tools & CLIs#community#python#trading
Community
TauricResearch/TradingAgentsPython
TauricResearch/TradingAgentsPython3.4kGitHub stars
Vercel Labs' Unix-like coding agent CLI written in Zig, shipping a typesafe_permission_reviewer builtin that runs the agent's permission decisions through Jev rather than an LLM call. Model-agnostic and embeddable via libfx and ACP.
Jev Tools & CLIs#community#coding-agents#cli
Community
vercel-labs/fxZig
vercel-labs/fxZig2.6kGitHub stars
Cisco's security scanner for agent skills added a System One analyzer: it sends the full skill text to POST /v1/systemone as one advisory Noul question and records the probability, while being deliberately unable to emit a finding or change a severity.
Jev Tools & CLIs#community#security#agent-skills
Community
cisco-ai-defense/skill-scannerPython
cisco-ai-defense/skill-scannerPython523GitHub stars
Classifies and splits PDFs, DOCX, and PPTX with local page text from LiteParse and category decisions from Jev.
Jev Tools & CLIs#community#python#document
Community
jerryjliu/docjevPython
jerryjliu/docjevPythonFrom the Jev guides
Original write-ups on what this entry is for.
The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.
Jev Tools & CLIs#gateway#vercel#cloudflare
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.
Jev Tools & CLIs#classification#email#security
Read article
From the Jev community
Posts from builders shipping with Jev right now.