Cua

// Overview

Cua is open-source infrastructure for computer-use agents: AI agents that see a screen, move the mouse, type on the keyboard, and run shell commands to finish tasks. It packages sandboxes (full VMs or containers) and SDKs behind one API, so an agent can drive a real desktop instead of just calling text APIs.

It targets developers building autonomous agents that need to operate software the way a person does. The same API works across Linux containers, Linux VMs, macOS, Windows, and Android, and it runs the same way on the cua.ai cloud or locally through QEMU, so you can prototype on your own machine and scale out later.

As an agent framework for computer and browser use, Cua sits between your agent logic and the operating system. Alongside the core sandbox SDK it ships related pieces: Cua Drivers for background desktop control, Cua-Bench for evaluation and RL environments, and Lume for macOS virtualization on Apple Silicon.

// What it does

• One Python API (Sandbox, Image) that works the same across Linux, macOS, Windows, and Android targets
• Runs in the cloud on cua.ai or locally with QEMU, including bring-your-own-image .qcow2 and .iso files
• Built-in actions for screenshots, mouse clicks, keyboard typing, shell commands, and multi-touch mobile gestures
• Cua Drivers control native desktop apps in the background without stealing cursor or focus, via a CLI and MCP server
• MCP integration so clients like Claude Code, Cursor, and Codex can drive a desktop
• Cua-Bench environments for evaluating agents on OSWorld, ScreenSpot, and Windows Arena, with trajectory export for training

// Getting started

Install the cua package and open an ephemeral sandbox to run your first computer-use actions. Requires Python 3.11 or later.

pip install cua

from cua import Sandbox, Image

async with Sandbox.ephemeral(Image.linux()) as sb:
    result = await sb.shell.run("echo hello")
    screenshot = await sb.screenshot()
    await sb.mouse.click(100, 200)
    await sb.keyboard.type("Hello from Cua!")

claude mcp add --transport stdio cua-driver -- cua-driver mcp

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Building an autonomous agent that completes multi-step GUI tasks across web and desktop apps
• Giving a coding agent (Claude Code, Cursor, Codex) a real computer to operate in the background via MCP
• Evaluating or training computer-use models on benchmarks like OSWorld and Windows Arena, then exporting trajectories
• Spinning up disposable macOS, Windows, or Android environments for automated testing on Apple Silicon or local QEMU

// How Cua compares

Cua alongside other open-source computer & browser use tools AI/TLDR tracks, ranked by GitHub stars.