cua
Cua gives AI agents computers they can use. Rather than being an agent itself, it provides the environment and controls: Cua Spaces gives agents full desktops (local macOS VMs on a Mac, or Linux images) on machines you own; Cua Driver lets agents inspect and operate native apps and browsers on macOS, Windows and Linux through a CLI, MCP or typed SDKs; Lume creates macOS and Linux VMs on Apple Silicon; and Cua Bench creates simulated tasks to evaluate agents and export trajectories.
It suits teams building or evaluating computer-use agents who need isolated, reproducible machines, and developers who want their coding agent (Claude Code, Codex, Cursor or OpenClaw) to operate desktop apps safely. Sandboxes can be created from the `cua` CLI or from SDKs in Python, TypeScript, Swift and Kotlin, and Cua Driver can work in the background without taking over your pointer. Spaces adds features such as moving a signed-in app into a sandbox and sharing a desktop between you and an agent.
Getting started: run `curl -fsSL https://cua.ai/install.sh | sh` on macOS or Linux (PowerShell on Windows). The installer sets up the `cua` CLI, the Spaces app on macOS 26+ and the Cua Driver MCP and skill; then `cua sb create ubuntu --name dev` creates a sandbox.
Features
Why choose it
Trade-offs
Maintenance check
- Last commit
- 2026-10-23
- Commits, last 90 days
- 100+
- Releases, last 90 days
- 30
- Latest release
- cua-spaces-v0.13.1 (2026-10-22)
- Open issues
- 1247
- Recent issues with at least one comment
- 1/12
Status rule: archived on GitHub = Archived; last commit within 90 days = Active; 91 to 365 days = Slowing; older = Stale. All figures come from the GitHub API on the date shown.
Most discussed open issues
- #353 An error because the Docker image trycua/cua-ubuntu:latest doesn't have a version compatible with arm64 arch (Apple Silicon Mac)13 comments · 2025-08-19
- #3550 [RFC]: Isolated background input on Hyprland12 comments · 2026-09-05
- #22 Document how to publish a lume VM image to a custom GHCR registry12 comments · 2025-02-08
- #3971 Feature request: configurable post-action observation and text-field settle delays in MCP10 comments · 2026-09-18
- #3796 [RFC]: cooperative cancellation of one in-flight driver call, with owned native cleanup10 comments · 2026-09-13
Issue titles are quoted from GitHub. Open the link to check whether a problem still applies to the version you use.
Who it fits
- You are building Computer-Use agents and need robust, self-hostable isolated desktop/VM sandboxes (Docker, QEMU, Apple Vz).
- You need native app operating drivers, standardized benchmarking harnesses, and secure isolation for autonomous agent task execution.
- You only need simple UI macros on your personal physical machine (Windows-MCP or local desktop helpers are easier).
- Your agents interact solely via text APIs, shell commands, or databases and have no need for a virtual graphical desktop.
How it differs
vs Windows-MCP★ 10.5kCUA provides sandboxed virtual desktops (QEMU/Docker) for safe agent testing; Windows-MCP controls the user's host physical Windows machine directly.
CUA simulates full operating system environments for any desktop app; Browserbase specializes in managed cloud browser instances strictly for web automation.
Compatibility
Quick start
Use cases
Alternatives
Comments
No comments yet. Be the first!