Organization repository on GitHub
Jev for Computer Use
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
CUA-JEV
Reference framework for Jev-driven computer use where a task adapter proposes legal action candidates, Jev picks one action and its execution route, and code guards, executes and verifies the result.
- Stars
- 19
- Language
- Python
- Published by
- Community
- Author
- ZJU-REAL
- Last verified
- Oct 2026
About this project
Webpage · Watch demos · 中文文档.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Jev picks the action and its channel in one call, so a step can go out as GUI input, DOM, MCP, CLI or a file API.
- ActionGuard checks observation freshness, candidate identity, path scope, write permissions and confirmation needs before dispatch.
- Four recorded Windows runs of 18-21 actions each, with per-case channel breakdowns and expandable decision traces.
- Earlier workflow timings show structured routes helping most on VS Code and Explorer, where GUI-only took 90.8 and 211.1 seconds.
Watch out
Desktop workflows and open-desktop depend on Windows UI Automation, Excel COM and Explorer; macOS and Linux desktop adapters are neither implemented nor validated, and the experimental browser path still needs a live macOS run. The recorded cases are described as single bounded successes, not repeated success-rate, speed or cost benchmarks. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Both. A text model reads live DOM and registered tool descriptions and proposes grounded next actions; Jev then selects one concrete action and channel at every step. The project uses the existing Jev API and trains neither Jev nor a new computer-use model, and no task-specific router is trained.
Single successful bounded research-to-editor runs, in the project's own wording, not repeated success-rate, speed or cost benchmarks. They use structured observations and a text planner with zero VLM calls, stay bounded to one browser origin plus registered tools, and Jev selected DOM in these recordings even where GUI routes were available.
The typed decision loop, guards and traces are designed to be portable, and CI is configured to run unit tests on Linux and macOS as well as Windows. But the published desktop workflows and open-desktop depend on Windows integrations, so macOS and Linux desktop adapters are unimplemented and unvalidated.
The framework narrows open-ended planning to constrained action selection: adapters propose real, already-typed actions, Jev chooses one online, and deterministic code handles safety and factual verification. The published comparisons are medians of available successful runs, with one pilot run per Codex workflow, so read them as illustrative.
ActionGuard checks observation freshness, candidate identity, path scope, write permissions and confirmation requirements before dispatching to a registered executor. Jev only selects a candidate ID from already-typed legal intent-and-channel options, so it never emits arbitrary shell text, and after execution an independent verifier checks the new application state and writes a JSONL trace rather than trusting the executor receipt.
A step can go out as GUI input, DOM, COM, MCP, CLI or a file API. In the four recorded Windows runs, each of 18 to 21 actions had one Jev call and one model call, and the selected channels broke down as DOM around 8 to 10, MCP 8, plus a couple of CLI and one file API each, with zero VLM calls. Jev selected DOM even where GUI routes were available.
In the v2 records, Jev Hybrid versus GUI Only versus Codex Hybrid: VS Code ran 14.0, 90.8 and 57.9 seconds, and Explorer 13.2, 211.1 and 50.9, so structured routes help most where GUI-only sequences get long. Edge was 79.5, 87.4 and 77.4, and Excel 84.2, 83.2 and 39.2, where Jev still chose GUI routes. Jev values are medians of available successful runs and Codex has one pilot run per workflow.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPython164GitHub stars
End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zwliJay/jev-forgePython
zwliJay/jev-forgePython63GitHub stars
Claude Code kit that guards PreToolUse calls with a free pre-filter plus one Jev question for ambiguous cases, and adds sub-agent sizing, rechecks, search, daemon speedups, and monitoring.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
jonathanavis96/jev-kitPython
jonathanavis96/jev-kitPython70GitHub stars
Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bladedevoff/stuntdPython
bladedevoff/stuntdPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.