Cubicles
Cubicles

Give Claude Code, Cursor or Codex a cloud computer with MCP

Your AI assistant already lives in your editor. With one MCP server it can also use a full Ubuntu computer in the cloud: run builds, test in a real browser and show you a screenshot.

Blog ·

The Model Context Protocol (MCP) is how AI apps like Claude Code, Cursor, VS Code, Codex and Claude Desktop connect to outside tools. Add an MCP server and your assistant gains new abilities, described in a way the model understands. New to MCP? Start with what MCP is and how it works.

This guide shows how to give your assistant a whole computer through MCP: a cloud desktop it can run commands on, read and write files on, and look at.

Why give your AI a remote computer?

  • Keep risky commands off your laptop. Let the assistant install packages, run unfamiliar scripts and try things on a machine you can wipe.
  • A clean Linux environment. Reproduce a bug on Ubuntu without a VM on your own machine.
  • Long jobs don't block you. Builds, data processing and test suites run on another machine while you keep working.
  • A real screen. The assistant can open an app or a browser on the remote desktop and take a screenshot to check what it did.
  • It keeps going. The computer stays on, so a job started from your laptop survives your laptop going to sleep.

What your assistant can do

The Cubicles MCP server gives your assistant these tools:

ToolWhat it does
list_desktops, get_desktopSee your computers and whether they're on
start_desktop, stop_desktop, restart_desktopTurn them on and off
run_commandRun any shell command as the computer's user, with sudo
read_file, write_fileRead and write text files
take_screenshotSee the screen
list_agents, get_agent, redeploy_agentCheck on agents running on your computers, and set one up again

Commands run with the desktop's display, so a graphical app started with run_command opens on the computer's screen, where take_screenshot can see it and you can watch it live.

Set it up in three steps

1. Create a computer

Sign up and create a computer. Your first one is free, with no card needed, and it's ready in about a minute.

2. Make an API key

In the app, open Developers and make an API key. Give it a name like “Claude Code on my laptop” and, ideally, an expiry date. It's shown once, so copy it somewhere safe.

3. Add the server to your AI app

Claude Code: run in a terminal
claude mcp add --transport http cubicles https://api.cubicles.cloud/api/v1/mcp \
  --header "Authorization: Bearer cub_…"
Cursor: add to ~/.cursor/mcp.json
{
  "mcpServers": {
    "cubicles": {
      "url": "https://api.cubicles.cloud/api/v1/mcp",
      "headers": { "Authorization": "Bearer cub_…" }
    }
  }
}

Replace cub_… with your key. Setup for VS Code, Codex CLI and Claude Desktop is in the MCP docs. The Developers page in the app shows every snippet with your new key filled in.

Things to ask your assistant

  • “Start my atlas computer, clone this repo there, install dependencies and run the test suite. Tell me what fails.”
  • “Run this migration script on my cloud computer, not here, and show me the output.”
  • “Open the app's sign-up page in Firefox on my cloud computer, take a screenshot and tell me if the layout is broken.”
  • “Check whether my OpenClaw agent is running, and set it up again if it isn't.”
  • “Write this config file to the cloud computer, restart the service and tail its log.”

How it works under the hood

  • Transport. Streamable HTTP at one URL. It's stateless, so there's no session to keep alive.
  • Auth. Your API key, sent as a bearer token. The key acts as you, and only on your computers.
  • Commands. run_command runs bash as the computer's user, who has passwordless sudo, and returns the exit code, output and errors.
  • Errors the model can fix. If a tool fails because of what it was asked, like a computer that's asleep, the reason comes back as text for the model to read and correct.
  • Limits. Commands, file reads and writes, and screenshots count towards 120 a minute per account.

The same actions are available as a REST API, if you'd rather script them. See the API reference.

Using it safely

  • Treat the key like a password. Anyone with it can run commands on your computers. Set an expiry and delete keys you no longer use.
  • One key per app. A separate key for each tool you connect means you can revoke one without breaking the others.
  • Keep secrets off the experiment box. Don't store production credentials on a computer your assistant uses for trial and error.
  • Review before it runs. Most AI apps ask before calling a tool. Leave that on for run_command until you trust the workflow.

Want the same from a script instead of an AI app? The REST API covers it, and more: see the Cubicles API.

MCP or an agent on the computer?

These are two different setups, and many people use both:

  • MCP: your assistant stays in your editor and reaches out to the computer when you ask. It works while you're working.
  • An agent on the computer: OpenClaw, Hermes, Claude Code or Codex lives on the computer itself and keeps working when you're gone, reachable by chat. See how to run an AI agent 24/7.

Ready to try it? Create your free computer, make a key, and ask your assistant to take its first screenshot.

Frequently asked questions

What is an MCP server?

MCP, the Model Context Protocol, is a standard way for AI apps to connect to outside tools. An MCP server offers tools, like running a command or taking a screenshot, that the AI can call when it needs them.

Which AI apps can connect to the Cubicles MCP server?

Any app that supports MCP over HTTP with a header for auth, including Claude Code, Cursor, VS Code and Codex CLI. Claude Desktop connects through the mcp-remote bridge. Setup for each is in the Cubicles MCP docs.

Can my AI see the remote computer's screen?

Yes. The take_screenshot tool returns the screen as an image the model can read, and you can watch the same screen live by opening the computer in your browser.

Is it safe to let my AI run commands on a cloud computer?

Safer than on your laptop, because the computer is separate from your files and network. Treat the API key like a password, set it to expire, and keep production secrets off that computer.

Does the remote computer stay on after my AI disconnects?

Yes. It stays on until you, or your assistant, put it to sleep. Its files are kept while it sleeps.

Give your agent a computer