CUA

Overview

CUA provides secure desktop sandboxes for computer-use agents. Instead of letting AI agents operate your real computer, CUA creates isolated desktop environments where agents can safely click, type, and navigate — perfect for testing and automation without security risks.

How It Works

1. Create sandbox — Spin up an isolated desktop environment
2. Connect agent — AI agent sees the desktop via screenshots
3. Execute tasks — Agent operates the desktop naturally
4. Destroy — Clean up when done

Use Cases

• Safe computer use — Let agents operate desktops without risk
• Software testing — AI-driven testing of desktop applications
• Process automation — Automate complex desktop workflows
• Training — Generate training data for computer-use models

Getting Started

from cua import DesktopSandbox

sandbox = DesktopSandbox(os="ubuntu")
sandbox.start()
screenshot = sandbox.screenshot()
sandbox.click(x=100, y=200)
sandbox.type("Hello from the sandbox!")

Example

Task: "Install Firefox and visit example.com"
Agent: Opens terminal → types 'apt install firefox' → 
opens Firefox → navigates to example.com → confirms
All in a disposable sandbox.

Alternatives

• UI-TARS Desktop — ByteDance’s multimodal agent
• E2B — Code execution sandboxes
• Anthropic Computer Use — Claude’s built-in capability