trycua/cua: one repository for agent desktops, a desktop driver, Apple Silicon VMs and a benchmark, under two licenses

A Space is a full VM desktop handed to an agent

Cua's pitch is short: give AI agents computers they can use. Cua Spaces is the desktop app that does it. Each desktop is a Space, either a macOS VM built locally on your Mac or a Linux or Omarchy image. Spaces run on your Mac, on other computers you own, and in your own cloud account on AWS, Google Cloud or Modal, and the app keeps them in the menu bar and the notch.

Three mechanisms set Spaces apart. Teleport moves a signed-in app such as Chrome or Slack into a Space, where it opens still signed in; the sessions sit in the Cua Keyvault, encrypted on your Mac, and only reach a Space after you approve. Multiplayer puts you and your agents on the same desktop with separate cursors, so you can step in for a decision and hand control back. And Spaces images ship with cua-spacesd, so an agent gets processes, files, screenshots and input as soon as a Space starts, without extra setup inside the VM.

The Spaces installer needs macOS 26, and its direct download links trail the releases

Spaces installs on macOS 26 or later with a piped script:

curl -fsSL https://cua.ai/install.sh | sh

On macOS that installer selects the Cua Spaces app by default and adds the `cua` CLI. Anyone on an older macOS release cannot use this route at all, and the minimum is a hard requirement rather than a recommendation.

The alternative routes are a signed universal `.dmg` and a `.pkg`, but both links on the front page point at the Cua Spaces 0.9.4 release. The release list has moved well past that: Cua Spaces v0.19.4 shipped on 2026-11-03, with v0.20.0-beta.1 tagged the same day. If you click the DMG link you install a build ten minor versions behind the current one. Pick the asset from the latest release page instead, and avoid the beta tag unless you want prerelease behavior on the machine that holds your Keyvault sessions.

MIT repository, FSL-1.1-MIT app, and per seat pricing

The repository is MIT licensed and the Python workspace declares MIT as well. The Spaces app is the exception: it is source-available under FSL-1.1-MIT, a different license from the rest of the code, and you should read the licensing section before treating it like the MIT libraries around it.

Pricing follows the same split. Spaces is free for individuals. Pro costs $18 a month, and Teams costs $30 per seat a month, available from one seat. Cua Driver, Lume, the SDK and Cua Bench do not carry those prices on the front page.

The consequence for a company is that the convenient path, the menu bar app with Teleport and Multiplayer, is the one with license and seat terms attached. A team that only needs agents to operate apps can stay on Cua Driver and Lume; a team that wants shared desktops with signed-in sessions is buying into the FSL terms and a monthly bill.

Cua Driver operates native apps in the background only where the platform allows

Cua Driver gives an agent tools to inspect and operate native desktop apps and browsers on macOS, Windows and Linux, reached through a CLI, MCP or typed SDKs. On macOS and Linux it installs with:

/bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)"

On Windows, from PowerShell:

irm https://cua.ai/driver/install.ps1 | iex

The headline mechanism is background delivery: an agent works without moving your pointer or taking focus. The qualifier matters. It applies only when the app and the platform support it, and the boundaries live on a separate platform support page. If you plan to keep working on the same machine while an agent drives other apps, check that page for your apps first, because outside those boundaries the agent and you compete for the same pointer.

The first exercise is concrete: connect your agent, ask it to compute 6 × 7 in Calculator, and have it verify that the app shows 42. Integrations exist for Claude Code, Codex, Cursor and OpenClaw. A demo shows two Driver sessions selecting cells in LibreOffice Calc and objects in Inkscape on an Omarchy desktop while a terminal stays in the foreground.

Lume is tied to Apple Silicon and Virtualization.Framework

Lume creates and manages local macOS and Linux VMs using Apple's Virtualization.Framework, and it runs on Apple Silicon only. It installs with:

/bin/bash -c "$(curl -fsSL https://cua.ai/lume/install.sh)"

The first tutorial creates a vanilla macOS Tahoe VM from an Apple restore image, starts it and connects over SSH, using the Lume CLI directly and explaining the unattended setup defaults along the way.

The limitation follows from the framework choice. Lume does not run on Intel Macs, Windows or Linux hosts, so local macOS VMs for agents are available only to people with Apple Silicon hardware. Teams on other hosts have to use Spaces images on their own cloud or other machines, or skip local macOS VMs entirely.

The Python workspace accepts only 3.12 and 3.13

The Python side is a uv workspace named cua-workspace with `requires-python = "<3.14,>=3.12"`. Python 3.11 and earlier are out, and so is 3.14. If your agent stack already moved to 3.14, the workspace will refuse to resolve until you keep a separate 3.12 or 3.13 environment for it.

Workspace members are `libs/python/agent`, `libs/python/core`, `libs/python/som` and `libs/python/bench-ui`, published as cua-agent, cua-core, cua-som and cua-bench-ui. cua-sandbox is deliberately not a member: consumers resolve it from `libs/python/cua-sandbox` as an editable path, and it depends on the cua SDK through a UniFFI binding. Top level dependencies are `openai>=2.20.0,<3.0.0` and `anthropic>=0.67.0`, and uv constraints pin fastmcp to `>=3.2.0,<3.3.0`.

Releases go through Release Please, which opens a release PR per package from conventional commit titles. Running `make show-versions` prints the released version of every package from `.release-please-manifest.json`. That per package model explains the release list: cua SDK 0.10.1 and Cua Spaces v0.19.4 both shipped on 2026-11-03 as separate tags.

Bring your own model, or use CUA-S1 for the decisions

Cua supplies the computer and the automation tools; the agent and model are yours. You can connect your own, or use CUA-S1, a set of small, specialized models for computer-use decisions. The project frames this as Computer-Use 2.0: an agent moving between code, APIs and graphical interfaces inside the same task, rather than driving only a screen.

Cua Bench covers evaluation. Its first exercise creates a simulated task and verifies it, and the project positions Cua Bench as the place for evaluating computer-use agents and for training and data generation at scale. The examples directory backs this up with folders for agents in sandboxes, streaming, training, Python and Driver usage.

One pattern runs through every install in the repository: Spaces, Driver and Lume all install through a script fetched from cua.ai and piped into a shell. On a machine that will also hold your signed-in sessions, download each script and read it before running it. The project is not archived, and its last push landed on 2026-11-02.

Editorial conclusion

trycua/cua fits developers who want their own agent to operate real desktops, especially on Apple Silicon Macs where Spaces and Lume run locally, and on macOS, Windows or Linux for Cua Driver. It is a weaker fit for teams that need everything under MIT, since the Spaces app is FSL-1.1-MIT and paid beyond individual use, or for anyone on an Intel Mac or on Python 3.14. Before adopting it, check the platform support page for Cua Driver background delivery, pick installers from the current release rather than the pinned 0.9.4 links, and read each piped install script before running it.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes