Give AI agents computers they can use.
Cua Spaces gives your agents full desktops on your Mac and on machines you own. This repository also holds Cua Driver for desktop automation, Lume for local VMs, CUA-S1 decision models and Cua Bench for evaluating computer-use agents.
[](#cua-spaces)
[](#cua-driver)
[](#lume)
[](#cua-bench)
Bring your own agent and model, or explore CUA-S1 for specialized decisions. Cua provides the computer and automation tools. Computer-Use 2.0 describes an agent moving between code, APIs, and graphical interfaces within the same task.
Cua Spaces
Cua Spaces is a desktop app that gives your agents full desktops. Each desktop is a Space: a macOS VM built locally on your Mac, or a Linux or Omarchy image. Spaces run on your Mac, on other machines you own, and in your own cloud account (AWS, Google Cloud or Modal), and the app keeps them in your menu bar and notch.
Teleport. Move a signed-in app, such as Chrome or Slack, into a Space and it opens there still signed in. Your sessions stay in the Cua Keyvault, encrypted on your Mac, and reach a Space only after you approve.
Multiplayer. You and your agents work on the same desktop, each with your own cursor. Step in to make a choice, then hand the desktop back.
Agent-ready images. Spaces images ship with cua-spacesd, so agents get processes, files, screenshots and input as soon as a Space starts.
Spaces is free for individuals. Pro is $18 a month and Teams is $30 per seat a month, available now from 1 seat. See pricing. The app is source-available under FSL-1.1-MIT.
Give your agent tools to inspect and operate native desktop apps and browsers on macOS, Windows, and Linux. Connect through the CLI, MCP, or typed SDKs. Background delivery lets agents work without moving your pointer or taking focus when the app and platform support it; see platform support for the boundaries.
Your first result: connect your agent, ask it to compute 6 × 7 in Calculator, and have it verify that the app displays 42. The tutorial covers platform setup, permissions, and agent connection.
Two Cua Driver sessions select cells in LibreOffice Calc and objects in Inkscape on an Omarchy desktop while a terminal stays in the foreground. Watch the 50-second demo.
Your first result: create a vanilla macOS Tahoe VM from an Apple restore image, start it, and connect over SSH. The tutorial uses the Lume CLI directly and explains the unattended setup defaults.
One SDK and one cua command for local VMs and containers, and for any machine that runs cua-spacesd.
macOS / Linux
curl -fsSL https://cua.ai/install.sh | sh
Windows (PowerShell)
irm https://cua.ai/install.ps1 | iex
In a terminal the script first says what Cua Spaces is. The command-line tools (the cua program) are always installed; you choose whether to add the Cua Spaces app (on by default) and Cua Driver, which lets AI agents use your apps while you keep working (off by default). It then runs cua auth login, which signs you in and offers to set up your AI coding agents (Claude Code, Codex, Cursor, and others). The app asks on its first run whether your other devices may use this computer. Add Cua Driver with sh -s -- --select cua-driver, or skip the choice with --only cua-driver. See the installer options.
cua sb create ubuntu --name dev # a gVisor container, set up on first use
cua sb exec dev uname -a
cua sb screenshot dev
cua sb rm dev
The same sandbox from Python:
from cua_sandbox import Image, Sandbox
async with Sandbox.ephemeral(Image.linux(), local=True) as sb:
print((await sb.shell.run("uname -a")).stdout)
SDK: the same API in Python (pip install cua), TypeScript (@trycua/cua), Swift (Cua) and Kotlin (generated bindings), running embedded in your process or through a shared cua daemon.
Sandboxes need no agent inside. Readiness comes from the runtime and optional port probes. Images that ship cua-spacesd (port 3211) add processes, files, screenshots, input through cua-driver, and low-latency video and audio streaming.
Your own machines.cua host setup makes this machine reachable through the cua.ai relay with no port forwarding.
CUA-S1 is our family of small, specialized System 1 models for computer use. We use "System 1" as an engineering analogy for fast, bounded decisions, such as choosing which value belongs in a field or whether to leave an element alone. It is not a strict classification of model architectures or a replacement for a general-purpose agent's planning and reasoning.
The first research profile focuses on forms: scoring decisions from structured interface elements and document values rather than generating a response token by token. Application code orders the actions, and the optional Cua Driver integration handles execution with explicit action boundaries.
The project includes Python model code, synthetic-data generation, training, and evaluation. The GitHub component is an early, source-only research release; model weights are hosted separately on Hugging Face. The source is MIT-licensed. Check each model and dataset card for its scope, limitations, and artifact-specific license.
Build computer-use tasks, evaluate agents, and export trajectories for training. Start with a simulated task that requires no VM, Docker, or model API key.
Your first result: create a small task, run its reference solution, and verify that its evaluator reports a reward of 1.0. Then try the same task yourself.
Cua no longer maintains these. They will not receive updates or fixes, and existing releases stay on npm and PyPI. See Migrate from deprecated packages.
Everything outside Cua Spaces and Cua Driver Bench is MIT-licensed (LICENSE): the cua SDK and its Python, TypeScript, Swift and Kotlin packages, the cua command and cua daemon, Cua Driver, Lume and the rest of this repository unless a directory says otherwise.
Cua Spaces is source-available under FSL-1.1-MIT: the Spaces apps, cua-spacesd, the Cua Keyvault, teleport, Cua Volume (cua-volume) and the streaming client, codecs and viewers. The streaming wire protocol stays MIT. It is free to use, self-host and build on, with no competing hosted service, and each release becomes MIT two years after it ships.
Cua Driver Bench (libs/cua-driver-bench) is under the same FSL-1.1-MIT; versions published before 11 Oct 2026 (#4973, #4978) were MIT.
The MIT parts never depend on the FSL parts. Some subdirectories carry their own licence; LICENSING.md lists each one and how the two fit together.
Offering Spaces, or Cua Driver Bench, as a hosted or managed service? See COMMERCIAL.md. For use of our names and logo, see TRADEMARKS.md.
cua-som is an optional package licensed under
AGPL-3.0-or-later. Its Ultralytics dependency retains its own license; inspect
the resolved dependency version and its notices before redistribution.
The Microsoft OmniParser repository states
CC-BY-4.0 for its
repository content. Model files downloaded from the separate OmniParser model
repository are distinct artifacts; verify the terms published with the exact
model revision before redistributing them.
The optional cua-perception extension is installed separately from the MIT
Cua Driver, from the signed assets of a cua-perception-v GitHub
release. Each release combines an AGPL-3.0-only OmniParser model artifact,
Apache-2.0 PP-OCR model artifacts, and a separately packaged ONNX Runtime.
The extension is not MIT licensed. Redistributing it, or offering it
to users over a network, can trigger AGPL-3.0 source obligations. Cua does
not relicense the detector and cannot grant other terms for it. Read the
perception third-party notices and precautions
before you install, redistribute, or host an extension artifact.
Trademarks
Apple, macOS, Ubuntu, Canonical, and Microsoft are trademarks of their respective owners. This project is not affiliated with or endorsed by these companies.
