CUA-JEV
Reference framework where Jev picks a concrete computer-use action and its execution route from legal, verifiable candidates.
Overview
CUA-JEV is an open-source reference framework for Jev-powered computer use. Task adapters turn structured state from the browser, desktop UI, office applications, terminal or filesystem into legal, executable, verifiable action candidates. Jev selects both the concrete action and which computer-use channel should perform it, for example a structured tool versus a physical GUI action; the framework guards and executes the choice, verifies the resulting state and observes again.
It includes four task-specific workflows and recorded model-plus-Jev Windows case studies, plus experimental open-task paths where a model plans and Jev selects among dynamically discovered DOM elements or UI Automation controls. The README is explicit about scope: Jev chooses among constrained options and does not interpret screenshots or write scripts, desktop adapters are Windows-only, and arbitrary-task generalisation is not claimed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate action should be executed next?(inferred) | — |
| Choice | Which execution route (structured tool or physical GUI) should perform it? | — |
How it works
- Adapter observes structured state and builds legal action candidates
- Jev selects an action and execution route
- Guards check the choice before execution
- Execute, verify the new state, and observe again
Strengths
- Actions restricted to legal, verifiable candidates
- Separates planning (LLM) from action selection (Jev)
- Traces and guards designed to be portable
Caveats
- Desktop workflows depend on Windows integrations
- Arbitrary-task generalisation explicitly not claimed
- Custom licence
Getting started
See the project webpage https://zjureal.com/CUA-JEV/ and docs/OPEN_TASKS.md for the experimental paths.At a glance
- Type
- App
- Status
- research
- Language
- Python
- Runtime
- desktop/browser automation (Windows)
- Has code
- yes
- Has tests
- yes
- License
- Custom license
Use cases and integrations
How curated lists describe it
The \[experimental open-task paths\](docs/OPEN\_TASKS.md) separate model planning from Jev's typed action selection. They discover browser DOM elements or Windows UI Automation controls dynamically and offer grounded structured-tool / physical-GUI alternatives without a site- or app-specific click sequence. A twelve-action browser-to-VS-Code research case now runs end to end with a model and Jev, but \*\*arbitrary-task generalization is not claimed\*\*.
logicrw/awesome-jev-projectsJev for Computer Use.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Which candidate action to take next and which channel should execute it.
Does Jev read screenshots?
No, it selects among structured candidates built by adapters.
Which platforms work?
Published desktop workflows need Windows; the browser path is more portable but macOS needs a live run.
Related projects
Python decision layer for browser and desktop automation: code observes, acts and verifies; the model makes one typed choice per step.
ProjectbetaPythonLets coding agents run whole goals in your signed-in Chrome on macOS in background tabs, asking Jev or Claude Haiku only for real choices.
ProjectbetaComputer-use platform (Cua Spaces, Driver, Lume) with CUA-S1 small decision models and a preview jev-use example for bounded actions.
ProjectbetaPythonLocal stdio MCP server that executes whole browser and macOS goals with a Jev-driven observe-decide-act loop.
MCP serverbetaTypeScriptChrome extension port of Jev Ultrafast that drives your current tab, with Jev picking the operation and element each step.
Browser extensionbetaTypeScriptBrowser Use's browser agent where Jev picks an operation and page element in one request and a small LLM only writes text to type.
LibrarybetaPython