Drive your cloud desktops from Claude: our MCP server is live

Drive your cloud desktops from Claude: our MCP server is live

Until now, controlling a Le Bureau desktop meant either the dashboard or a curl against the API. It works, but it pulls you out of whatever you were doing — switch window, look up the desktop ID, build the request. A lot of friction for something you do ten times a day.

As of April 23, @lebureau/mcp is published on npm. Your desktops show up as tools that Claude Desktop, Claude Code, Cursor — or any Model Context Protocol client — can call directly. You describe what you want in plain English, the assistant picks the right tool and runs it for you.

What MCP is, briefly

The Model Context Protocol is an open standard Anthropic released last year. It's a stdio channel between a client (Claude Desktop, Claude Code, Cursor, and so on) and a local tool server: the server advertises a list of typed tools, the client surfaces them to the model, and the model invokes them when the conversation calls for it. No bespoke plugin, no API for the user to learn.

For us, this means we don't have to ship a separate integration for every client. We publish one MCP server and the whole ecosystem benefits. More importantly, it means we don't ask our users to leave the assistant they already live in to go click around somewhere else.

What it actually feels like

Once the server is wired up, you can type things like:

• "List my cloud desktops, tell me which ones are running."
• "Start the market-research desktop, open Chrome, send me a screenshot of Cursor's homepage."
• "Run this Python script on the data-pipeline desktop and come back with the logs."

You don't name the tool. Claude looks at the list the server exposes — desktops_list, desktop_chat, desktop_run_python, desktop_screenshot, a dozen more — and figures out which one fits. The assistant turns intent into an API call. You stay in the conversation.

The new thing isn't the technical capability — all of this was already possible through the REST API. The new thing is friction. Asking Claude "look at what's running and shut down the ones I haven't touched since yesterday" takes ten seconds instead of three minutes.

How it stays safe: per-key capabilities

Giving an assistant the power to run Python on your machines deserves a real guardrail. That's what the new capabilities system is for, shipped alongside the MCP server.

Every API key carries an explicit list of what it can do. Read-only on desktops? Keyboard and mouse control? Shell command execution? Knowledge base reads? You tick the boxes. Anything you don't tick gets refused, full stop.

The MCP server reads those capabilities at startup and filters the tool list it advertises to Claude accordingly. If your key only has desktop:read and desktop:chat, the desktop_run_python tool doesn't even appear in the catalog Claude sees. The model can't call what it can't see.

In practice, this gets you two things. First, you can mint several keys with different scopes — one for routine read-only automation, one for sessions where you let the assistant act, one for CI deploys. Second, if a key leaks or someone gets hold of your claude_desktop_config.json, the narrow scope keeps the blast radius small.

By default, the key creation form at /settings?tab=apiKeys pre-checks desktop:read and desktop:chat. That's the minimum useful surface to wire Claude Desktop up without surprises. Everything else — Python execution, mouse control, knowledge base writes — is opt-in, capability by capability. Each scope is documented in the capabilities reference.

Wiring it up in three steps

1. Mint a key. Go to /settings?tab=apiKeys, tick the capabilities you need (start with desktop:read and desktop:chat if you're feeling things out), copy the sk_live_... key that appears. It's only shown once.
2. Drop the config into your client. For Claude Desktop, open claude_desktop_config.json and add:

{
  "mcpServers": {
    "lebureau": {
      "command": "npx",
      "args": ["-y", "@lebureau/mcp@latest"],
      "env": {
        "LEBUREAU_API_KEY": "sk_live_..."
      }
    }
  }
}

For Claude Code, same block in .mcp.json at the root of your project. The npx call is launched by the client itself when you open a conversation — you don't run anything by hand.

1. Restart the client. On the next launch, the Le Bureau tools appear. Ask "list my desktops" to confirm it's talking.

The full walkthrough (plus the Cursor config) lives in the install guide.

What v0.1.0 deliberately doesn't do

This first cut is deliberately narrow. We'd rather ship a tight surface that we trust than a roomy one we don't.

• No creating or destroying desktops via MCP. Starting an existing desktop, sure. Provisioning a new one or wiping one, that stays on the dashboard. Cost and irreversibility deserve a deliberate click, not a tool call buried in a conversation.
• No admin tools. Anything that touches account settings, billing, or key management stays in the UI.
• No live event streaming. If you want to watch a long-running process unfold, open the desktop viewer in the dashboard. The MCP tools return results in one shot.

These are on the roadmap, with no promised dates. The current surface is what's useful without being scary.

If you want to try it

If Claude Desktop, Claude Code or Cursor is already where you spend your day, this is a five-minute thing to plug in. The MCP overview docs list the available tools and walk through a few recipes — kicking off a remote agent, having a desktop screenshot a page, pulling a file from the knowledge base to start a conversation with real context.

And if there's a tool you wish was there and isn't, tell us. This first set reflects what we saw ourselves reaching for most often, but we're very interested in the workflows we didn't anticipate.

---

Le Bureau provides persistent cloud desktops for AI agents. Full Linux environment, remote access, real-time supervision. Try it free during the beta.