# Cua Cua (https://cua.ai, `trycua` on GitHub) gives AI agents computers they can use: full desktops, with apps, a browser and a shell, that an agent drives and that you can watch. Its pitch, from the Spaces SDK launch on 19 October 2026: the big labs now give their bots their own computers ([[Claude Duet]], [[Dots (AI)|Dots]], [[Meta Muse]]), but none of them let you bring your own. Cua does. You run [[Claude Code]], [[OpenAI Codex|Codex]] and others on compute you choose, and you can spread jobs over several machines at once. Cua doesn't rent computers. You bring them. ## Spaces A Space is one desktop on one of your computers, set up for a task. It can be: - **A macOS VM on your Mac**, through Lume (needs macOS 26+ on Apple silicon) - **A Linux container** (Docker, Podman or Colima) or a Linux VM. There are [[Omarchy]] images too - **A machine you own**, like a Mac mini or a Linux box, reached through a relay, SSH or [[Tailscale]] - **A VM in your own cloud account**: AWS, Google Cloud and Azure (no stored key: short-lived tokens through a CloudFormation stack, a workload identity pool or a managed identity), or [[Hetzner]] and Modal (API token encrypted, Cua keeps only the ciphertext). Your cloud bills you; Cua stops creating machines when your monthly budget is hit Windows Spaces are a preview download. Things I find well thought out: - **Shared desktop, separate cursors.** Cua's own remote desktop protocol (RCDP) streams the Space over QUIC and handles input apart from video. You and each agent get your own cursor, and you can step in to take over - **Teleport.** Move an app session (tabs, profile, sign-ins) from your Mac into a Space as a bundle. Nothing is read before you approve it, sensitive items need Touch ID, and the Space checks the bundle's SHA-256 before opening it. That solves the "the agent needs to be logged in" problem without handing it your passwords - **Ready to use.** Permissions are granted and tools installed in advance. You connect a provider once and every Space can use it ## How agents use it `cua connect setup --agents claude,codex --yes` installs the `cua` skills and MCP server into each agent (or `claude mcp add cua -- cua daemon mcp` by hand). The [[Model Context Protocol (MCP)|MCP]] server exposes 33 Spaces tools plus sandbox and computer tools: create or reuse a Space, run bash in it, send a file, get a stream URL, call desktop tools, start and message agents inside it, share the Space, delete it. `--permissions spaces:readonly` limits an agent to looking. Supported agents: Claude Code, Codex, [[Cursor]], [[Gemini CLI]], Cline, Kiro and [[OpenClaw]]. From code, the Spaces SDK (Python `pip install cua`, TypeScript `@trycua/cua`, plus Swift and Rust) does the same: ```python spaces = cua.embedded().spaces() created = await spaces.create(cua.SpaceCreateOptions()) space = await spaces.space(created.space.id) out = await space.bash("uname -sm", None) shot = await space.screenshot(None) await spaces.delete(created.space.id) ``` `SpaceCreateOptions` takes a `kind`: `container`, `vm` or `microvm`. The default image is `ghcr.io/trycua/linux:24.04`. ## The other pieces - **Cua Driver** (MIT): lets an agent inspect and operate native desktop apps and browsers on macOS, Windows and Linux, through a CLI, MCP or typed SDKs. This is the computer-use layer inside a Space - **Lume** (MIT): local macOS and Linux VMs on Apple silicon, on Apple's Virtualization framework - **CUA-S1**: small models specialized in computer-use decisions, starting with forms - **Cua Bench**: build computer-use tasks, evaluate agents, export trajectories The SDK has no agent loop of its own. You bring the harness; Cua brings the computer. ## Pricing and license - Free: Spaces on all your computers, Teleport, sharing, up to 2 live sessions and 20 remote hours a month - Pro: $18 per month ($15 yearly), 5 sessions, 200 hours - Teams: $30 per seat per month ($25 yearly), shared team computers, roles, device approval Cua Spaces is source-available under the [[Functional Source License (FSL)|FSL-1.1-MIT]]: each release becomes MIT two years later. Cua Driver and Lume are MIT. The repository has about 31.7k stars. Anonymous telemetry is on by default (`cua telemetry off` or `DO_NOT_TRACK=1`). ## How it compares - **vs [[Docker Sandboxes]], [[OpenSandbox]], [[Cloudflare Sandbox SDK]], [[Vercel Sandboxes]]**: those isolate code execution, mostly headless. Cua's unit is a full desktop with GUI apps that you can watch and share, and it runs on your own hardware or cloud - **vs [[GitHub Copilot Sandboxing]]**: Copilot limits what an agent's commands can touch on your machine. Cua moves the agent to another machine entirely - **vs [[Claude Duet]], [[Dots (AI)|Dots]]**: those come with a computer the lab hosts, for its own agent. Cua is the bring-your-own version, for any agent, and you keep the machine ## My take The "bring your own computer" angle is the right one. I don't want my agents' sessions, logins and files living on a lab's machine I can't inspect, and I don't want to be locked into one lab's agent to get a desktop. Cua puts that layer under my control, with agents I already use. The Omarchy images are a nice touch for me, since that's what I run. The parts I'd check before using it seriously: the relay and the dashboard are Cua's (the compute isn't), telemetry is on by default, the Spaces app is FSL rather than open source, and computer use is still slow and brittle whatever the sandbox. For most coding work, a container and a CLI are enough. Cua becomes interesting when the agent needs a real GUI app or a logged-in browser. ## References - Website: https://cua.ai - Spaces SDK announcement on X (2026-10-19): https://x.com/trycua/status/2112228154651881479 - GitHub: https://github.com/trycua/cua - Docs, Spaces: https://cua.ai/docs/spaces - Docs, use Spaces from code: https://cua.ai/docs/spaces/guides/create-a-space-with-the-sdk - Docs, connect your agent: https://cua.ai/docs/spaces/guides/connect-your-agent - Docs, connect a cloud: https://cua.ai/docs/spaces/guides/connect-a-cloud - Pricing: https://cua.ai/pricing ## Related - [[Claude Code Computer Use]] - [[Claude Computer use]] - [[Docker Sandboxes]] - [[OpenSandbox]] - [[Cloudflare Sandbox SDK]] - [[Vercel Sandboxes]] - [[GitHub Copilot Sandboxing]] - [[Claude Duet]] - [[Dots (AI)]] - [[Claude Managed Agents]] - [[Model Context Protocol (MCP)]] - [[Functional Source License (FSL)]] - [[Omarchy]] - [[AI Agent Permissions]]