reflex-cu
Computer-use MCP server for macOS and Windows that reads the UI tree and uses Jev to find elements, check screens and wait for states.
Overview
reflex-cu is an open-source computer-use MCP server for Claude Code, Codex and other agents. It controls local macOS and Windows desktops, or a remote Windows machine over SSH, by reading the accessibility or UI Automation tree (names, types, positions, states) rather than taking a screenshot every step, falling back to OCR when a window exposes nothing. The main README is in Chinese with an English version.
The 'reflex' idea is to hand frequent, low-thought judgments to a fast decision model instead of the agent: three tools use Jev (about 0.3 s per call). find takes a one-sentence description and returns the best-matching element with confidence, check asks a yes/no question about the current screen, and wait polls until such a question holds or times out. A steps tool runs a sequence and stops when Jev is unsure or a step fails. Jev can be reached through TypeSafe, OpenRouter or Command Code keys, or any compatible gateway; without a key, all other tools still work. Jev sees only text, so icon-only or game content still needs screenshots.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which UI element best matches this description? | — |
| Yes / no | Did the action succeed / is the screen in this state? | — |
| Yes / no | Has loading finished? | — |
How it works
- observe reads the UI tree (or OCR) into numbered elements
- find asks Jev which element matches a description and returns confidence
- Click or type; check asks Jev a yes/no question about the result
- wait polls until the condition holds; steps stops when Jev is unsure
Strengths
- Fewer screenshots for the agent to read
- Fast judge at about 0.3 s per call per the README
- Works with three key providers and custom gateways
- Remote Windows control over SSH
Caveats
- Screen text is sent to Jev's provider
- Jev cannot judge icon-only or game visuals
- Bring your own key and costs
Getting started
Install with Python 3.10+ (pip install -r requirements-macos.txt or requirements-windows.txt), set TYPESAFE_API_KEY or OPENROUTER_API_KEY, and register the MCP server with your agent; INSTALL_FOR_AI.md lets an agent do it.At a glance
- Type
- MCP server
- Status
- beta
- Language
- Python
- Runtime
- MCP server
- Has code
- yes
Use cases and integrations
How curated lists describe it
Computer use MCP server for Claude Code, Codex and other AI agents. Reads the UI tree instead of taking a screenshot every step, with Jev (TypeSafe System One) as a fast judge. Local macOS, local Windows, or remote Windows over SSH.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Which element matches a description, whether a yes/no condition holds on screen, and when to stop waiting.
Can I use it without a Jev key?
Yes; only find, check and wait need a key.
Do I need a TypeSafe account?
No; an OpenRouter or Command Code key also works.
Related projects
macOS computer-use package: a background app driver for any agent, plus a loop where Jev picks each UI action for a bounded subtask.
ProjectbetaPythonWindows MCP server and CLI where Jev observes windows and chooses controls so agents can hand off whole desktop tasks.
MCP serverbetaPythonAccessibility-first macOS computer-use MCP server and CLI, with optional Jev guards on targets and inputs before irreversible actions.
PluginbetaPythonWindows automation and durable research MCP server for Codex and Claude Code where Jev chooses among bounded UI Automation actions.
MCP serverbetaCodex computer-use skill where Jev picks the next on-screen element, action, completion and risk from text candidates.
Agent skillprototypeJavaScriptRust MCP server for desktop control via accessibility trees, with an optional decision model (Jev or compatible) for small choices.
MCP serverproductionRust