What is Cua?

Cua is an open-source platform for building, benchmarking, and deploying agents that can use any computer, with isolated, self-hostable sandboxes (Docker, QEMU, Apple Vz).

The Cua Stack

Cua consists of three main components:

1. Desktop Sandboxes

Isolated virtual environments where your agents can safely execute tasks:

• Cloud Sandboxes - Managed Linux, Windows, and macOS environments hosted by Cua
• Local Sandboxes - Docker containers, QEMU VMs, macOS VMs via Lume, or Windows Sandbox on your own machine

2. Computer Framework

A unified SDK for controlling desktop environments programmatically:

• Take screenshots and observe the screen
• Simulate mouse clicks, movements, and scrolling
• Type text and press keyboard shortcuts
• Run code and shell commands
• Works identically across all sandbox types

3. Agent Framework

Build agents that see screens, click buttons, and complete tasks autonomously:

• 100+ vision-language model options through Cua VLM Router or direct provider access
• Pre-built agent loops optimized for computer-use tasks
• Composable architecture for combining grounding and planning models
• Built-in telemetry for monitoring agent performance

How Computer-Use Agents Work

Computer-use agents operate through a continuous loop:

┌─────────────────────────────────────────┐
│  1. OBSERVE                             │
│     Take a screenshot of the screen     │
└──────────────────┬──────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────┐
│  2. UNDERSTAND                          │
│     Vision-language model analyzes      │
│     the screenshot and current goal     │
└──────────────────┬──────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────┐
│  3. DECIDE                              │
│     Determine the next action:          │
│     click, type, scroll, etc.           │
└──────────────────┬──────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────┐
│  4. ACT                                 │
│     Execute the action on the computer  │
└──────────────────┬──────────────────────┘
                   │
                   ▼
              Loop back to 1

This cycle repeats until the agent completes its goal or determines it cannot proceed.

Sandbox Options

Why Cua?

• Secure execution - Run AI coding assistants and computer-use agents in sandboxed environments
• Self-hostable - Deploy locally with Docker, QEMU, or Apple Virtualization
• Cross-platform - Same API across Linux, Windows, and macOS sandboxes
• Visual understanding - Agents see screens, adapt to UI changes, and complete complex tasks

Use Cases

• AI coding assistants - Isolated code execution environments for Claude Code, Codex CLI, OpenCode, and other AI coding tools
• Computer-use agents - Build agents that interact with any desktop application autonomously
• Workflow automation - Automate repetitive tasks across any application
• Testing - Run end-to-end tests that interact with real UIs
• Benchmarks - Evaluate agents on OSWorld, ScreenSpot, and other standardized tasks
• Research - Build, evaluate, and train computer-use AI agents

Getting Started

Ready to build your first agent? Continue to the Quickstart to get a computer-use agent running.