jev-windows-agent
Windows UI Automation backend for arc-cua that lets a planner hand off desktop subtasks to a Jev-driven click-and-type loop.
Overview
jev-windows-agent extends arc-cua, a fast action layer for computer-use agents, with a Windows UI Automation backend. Instead of screenshot-and-guess, it reads structured UIA elements from real Windows apps such as Notepad, Settings, Spotify and File Explorer and lets a Jev-driven decision loop click, type and navigate. A planner model (GPT, Claude, Gemini, DeepSeek or any model via OpenRouter) breaks a request into bounded subtasks with a goal, inputs, verification conditions and constraints, then hands each subtask to the loop so no frontier model is called for every click.
The demo shows a one-step plan to open Apple Music and play liked songs, completed in four actions with each decision's action, target, confidence and top candidates rendered live. The fork adds the Windows perception and execution backend, OpenRouter transport and optional confidence gating while keeping the original runtime, validation and Jev decision policy unchanged. Architecture, walkthrough and porting notes are documented, and the original project and macOS backend are credited to their authors.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which action and UI element should be used next to advance this subtask? | — |
How it works
- Planner splits the request into bounded subtasks
- Windows UIA backend observes structured elements
- Jev chooses action and target each step, optionally confidence-gated
- Stop on verification success or failure
Strengths
- No frontier-model call per click
- Structured UIA perception instead of screenshots
- Works with many planner models via OpenRouter
Caveats
- Windows-only backend
- Builds on another project's runtime
- Demo-level evaluation
Getting started
jev plan --yes --request "Open Apple Music and play my liked songs"At a glance
- Type
- Plugin
- Status
- prototype
- Language
- Python
- Runtime
- Windows desktop agent
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Windows UI Automation extension of arc-cua: a fast, JEV-powered decision loop for desktop computer-use agents
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev do here?
It picks each UI action and target inside a bounded desktop subtask.
What does the planner model do?
It breaks the request into subtasks with goals, inputs, verification and constraints.
Is it a new project?
It extends arc-cua with a Windows UI Automation backend.
Related projects
Python decision layer for browser and desktop automation: code observes, acts and verifies; the model makes one typed choice per step.
ProjectbetaPythonDurable Android sub-agent: a SQLite-backed worker observes the UI and Jev picks among valid actions, with verification and escalation.
MCP serverbetaPythonGo macOS controller where Jev picks the next accessibility action, the target element and whether the goal is done.
ProjectprototypeGoReference framework where Jev picks a concrete computer-use action and its execution route from legal, verifiable candidates.
AppresearchPythonmacOS voice agent where Jev picks each step's operation and target control, escalating to an LLM agent when unsure.
ArticleproductionAutoGPT agent platform, listed as adding Jev decision blocks for typed routing, filtering, scoring and confidence-gated next actions.
ProjectproductionPython