CtrlK
BlogDocsLog inGet started
Tessl Logo

cua-sandboxes

Create, use and clean up cua sandboxes (disposable Linux or macOS computers) locally or in the Cua cloud with the `cua` CLI or the cua SDK, and browse the web inside one. Use when a task needs an isolated machine to run code, test an app, drive a desktop GUI, browse or test a website, fill a web form, take a web page screenshot, or reproduce something without touching the user's own computer.

78

Quality

98%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

SecuritybySnyk

Low

Low-risk findings worth noting

SKILL.md
Quality
Evals
Security

cua sandboxes

A sandbox is a disposable computer: a container or VM on this machine or in the Cua cloud. The cua CLI and the cua SDK (Python, TypeScript, Swift, Kotlin, Rust) share one runtime and one list of sandboxes.

Before you start

cua --version            # the Rust cua CLI
cua auth status          # the cloud needs `cua auth login` (or CUA_CLIENT_ID/CUA_CLIENT_SECRET)
cua config get default.on   # where sandboxes run without --on (local unless changed)
cua runtime doctor       # local runtimes (containers, gVisor, QEMU, Lume); read-only

Local sandboxes need no account. Cloud sandboxes are metered: always delete them when done.

Create

cua sb create linux --name dev                 # local (default): a gVisor container
cua sb create linux --kind vm --name dev       # local Linux VM (QEMU)
cua sb create macos --name mac                 # local macOS VM (Apple silicon, Lume)
cua sb create linux --on cloud --name dev      # the Cua cloud
cua sb create ghcr.io/org/image:tag --name x   # any OCI image, local or --on cloud
cua sb create --on direct:127.0.0.1:3211 --token TOKEN --name mine   # an existing machine running cua-spacesd

--on is where (local, cloud, direct:<addr>), --kind what (auto, container, vm), --runtime which engine (auto; local gvisor, runc, qemu, lume; cloud gvisor, kubevirt). An impossible combination fails with invalid placement and lists the valid values. Local runtimes are set up on first use (cua runtime setup does it explicitly). Omit --name for a generated name. cua sb ls prints each sandbox's ref (local:dev, cloud:dev, direct:host:port); every NAME argument takes a ref or a bare name that is unique across locations (--local / --cloud narrow it). Add --json to any command for machine-readable output. exec, shell, cp, screenshots and GUI control go through cua-spacesd, so use an image that ships it (the linux alias does). Plain images still get lifecycle, logs and port-forward.

Use

cua sb ls                          # every location (--local or --cloud filters)
cua sb exec dev -- uname -a        # run a command
cua sb shell dev                   # interactive shell
cua sb screenshot dev              # save a screenshot
cua sb vnc dev                     # open the desktop in a browser

For GUI work, select the sandbox once and use cua do (see the gui-automation skill):

cua do switch dev
cua do screenshot
cua do click 400 300

Clean up

cua sb rm dev --force              # delete
cua fleet pools gc                 # remove idle auto-managed cloud capacity
cua sb suspend dev                 # or pause it and `cua sb resume dev` later

From code (Python)

import asyncio, cua

async def main():
    c = cua.embedded()
    sb = await c.sandboxes().create(cua.SandboxCreateOptions(
        on="local",
        image="ghcr.io/trycua/linux:24.04", name="dev"))
    env = await sb.spacesd(None)
    out = await env.run(cua.SpacesdCommand(program="uname", args=["-a"]))
    print(out.stdout.decode())
    await sb.delete()

asyncio.run(main())

The TypeScript, Swift and Kotlin bindings expose the same objects (Cua, sandboxes(), Sandbox.spacesd()).

From the cua MCP server

When the cua MCP server is connected, prefer its tools: images_list (which images exist), sandbox_list, sandbox_create, sandbox_get, sandbox_delete, then the computer_* tools (screenshot, click, type, shell, file) against the sandbox.

Browse the web

Browse inside a sandbox, never in the user's own browser. Everything below is cua MCP tool calls; no other browser tool or setup is needed.

  1. Pick an image: images_list {"browser": true} (the canonical ghcr.io/trycua/linux:24.04 ships Chromium).
  2. Create it with a browser: sandbox_create {"browser": true, "url": "https://example.com"} (add "on": "cloud" for the cloud). It returns id, session and browser.target_id / browser.tab_id.
  3. Drive the browser with call_tool {"space": <id>, "tool": <tool>, "arguments": {"session", "target_id", "tab_id", ...}}:
    • read: get_browser_state {"snapshot_format": "semantic_v2"} returns an outline and refs (p1:0); add "include_screenshot": true for an image;
    • browser_navigate {"url"}, browser_click {"ref", "delivery_mode": "foreground"}, browser_type {"ref", "text", "replace": true} (text "\n" with "mode": "keystrokes" presses Enter), browser_pointer (hover, scroll, drag);
    • read again after every page change: refs are per snapshot. To test a local app, run it in the sandbox (space_write, space_bash) and open http://127.0.0.1:<port>.
  4. Logged-in sites: only when the user asks. teleport_browser_session with the exact sites first returns what would move and a consent requirement; show it to the user and continue only after they approve. Never move a session without that approval.
  5. Clean up: sandbox_delete {"name": <id>}.

From a terminal: cua images ls --browser, cua sb create --browser --open https://example.com, cua sb view <id> to watch.

Rules

  • Never run destructive commands on the user's own machine when a sandbox would do.
  • Delete cloud sandboxes you created as soon as the task ends.
  • Report the sandbox name you used so the user can inspect or remove it.
Repository
trycua/cua
Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.