Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/trycua/cua/llms.txt

Use this file to discover all available pages before exploring further.

Cua is an open-source, MIT-licensed platform (github.com/trycua/cua) for Computer-Use 2.0 automation — letting AI agents operate computers by clicking, typing, reading screens, and running code. Unlike traditional RPA or browser automation tools, Cua treats the graphical interface as one tool surface in an agent loop, alongside code execution, shell commands, and structured API calls. The platform ships four main components: Cua Driver, Cua Sandbox, Lume, and Cua Bench.

​
The four components

Cua Driver

Drive native desktop apps in the background on macOS, Windows, and Linux. Agents click, type, and verify without stealing the cursor or keyboard focus. Connects to Claude Code, Cursor, Codex, and other harnesses via MCP over stdio.

Cua Sandbox

Create fully isolated, agent-ready computers — Linux containers, macOS VMs, Windows VMs, or Android — from a single Python API. One async with block starts and tears down the sandbox automatically.

Lume

Create and manage macOS and Linux VMs with near-native performance on Apple Silicon using Apple’s Virtualization.framework. Powers the local macOS sandbox backend inside Cua Sandbox.

Cua Bench

Evaluate computer-use agents on OSWorld, ScreenSpot, Windows Arena, and custom tasks. Export trajectories for training and reinforcement learning.

​
Computer-Use 2.0

The term computer use originally described agents that interpreted screenshots and controlled a GUI with mouse and keyboard events. Cua uses Computer-Use 2.0 to describe a broader model: an agent chooses among three complementary action surfaces during a single task.
  • Coding surface — the agent writes and runs code as the action itself.
  • Tool-use surface — the agent makes typed calls to tools, APIs, and MCP servers.
  • UI automation surface — the agent clicks, types, and scrolls through interfaces designed for humans.
All three surfaces share one filesystem and OS state inside a Cua Sandbox, or they operate against a real machine through Cua Driver.

​
Language support

Python SDK

pip install cua — supports Python 3.12 and 3.13. Includes the Sandbox SDK, Agent SDK, and CLI. The meta-package pulls in cua-sandbox, cua-agent, and cua-cli.

TypeScript SDK

Generated UniFFI bindings for TypeScript applications. Import @trycua/cua-driver to embed the Cua Driver native runtime directly in a Node.js or Electron app.
The core runtime is written in Rust. Language SDKs for Python and TypeScript are generated from the same typed contract through UniFFI bindings, so they share identical behavior across platforms.

​
Packages

PackageDescription
cuaUnified SDK meta-package (Sandbox + Agent + CLI)
cua-sandboxSDK for creating and controlling sandboxes
cua-agentAI agent framework for computer-use tasks
cua-driverBackground computer-use for macOS, Windows, Linux
cua-benchBenchmarks and RL environments for computer-use
lumemacOS/Linux VM management on Apple Silicon

​
Choose your path

Drive an existing app

Already have a machine with apps installed? Start with Cua Driver to give an agent access to it without disrupting your desktop.

Start a fresh sandbox

Need a clean, isolated Linux desktop? Spin one up with Docker in a single Python script — no VM setup required.

Create a macOS VM

Need a full macOS virtual machine on Apple Silicon? Lume downloads and configures one from an Apple restore image.

Benchmark an agent

Want to evaluate or train a computer-use model? Cua Bench provides verifiable tasks, reproducible environments, and trajectory export.

​
License and community

Cua is released under the MIT License. Source is at github.com/trycua/cua. Join the community on Discord or follow @trycua on X for updates.

Build docs developers (and LLMs) love

Get started for freeTalk to us