Orgo MCP for AI agents: control cloud computers from your MCP client

TL;DR

MCP for AI agents means giving an AI app a set of tools it can call mid-conversation through the Model Context Protocol. Orgo's MCP server gives the model tools for cloud computers: from your MCP client it can create a computer, click, type, run shell commands, and read the screen back (MCP guide). Add https://www.orgo.ai/mcp to a client that supports remote MCP servers and sign in, or run npx -y orgo-mcp-server locally with an API key. The tools are generated from Orgo's published API spec, so they follow the API as it changes. Orgo is not an agent: MCP is one of three ways to reach the same computers, next to the HTTP API and the CLI (Introduction).

What is MCP for AI agents?

The Model Context Protocol is an open protocol that lets an AI app, the client, connect to servers that offer tools. The client lists each server's tools and their arguments, and the model decides when to call one. For agent builders, that means you add a product's server to your client once, and the model can use its tools without you writing a wrapper for each product.

Many MCP servers wrap actions in a SaaS product, such as filing a ticket or searching a wiki. Orgo's tools cover workspaces, creating, starting and stopping computers, screenshots, mouse and keyboard, bash and Python, screens, files, and templates (MCP guide). The agent gets a full desktop that keeps its files across stops and restarts, not a single API call.

What is the Orgo MCP server?

• It ships two ways: as the npm package orgo-mcp-server, which a client starts as a local process, and as a hosted server at https://www.orgo.ai/mcp, which a client connects to by URL.
• Its tools are generated from Orgo's published OpenAPI spec. Descriptions, types, enums and required fields are read from the spec, so the tools match the HTTP API (MCP guide).
• There is one tool per API operation, named after its operationId: createComputer becomes orgo_create_computer.
• Tools are grouped by the API reference's own tags, such as Workspaces, Computers, Computer Actions, Screens, Files and Templates.
• A hosted sign-in sees only the tools it can use. It can manage computers, screens and files in the workspace you picked and read that workspace. Templates and workspace changes need an API key.

Orgo computers are full desktops with a browser and the standard userland, reachable over HTTP and VNC. A new computer is online in seconds, and files, installed software and configuration on disk survive stops and restarts (Introduction). MCP does not change what a computer is. It changes how your agent's client reaches it.

How do I connect with a sign-in?

For a client that supports remote MCP servers, add https://www.orgo.ai/mcp. The client opens Orgo's sign-in and consent page, you choose the workspace to authorize, and you approve. No API key is involved. In Claude Code:

claude mcp add --transport http orgo https://www.orgo.ai/mcp

Then run /mcp in Claude Code and finish the sign-in in your browser (Claude Code MCP docs).

The sign-in uses the OAuth authorization code flow with PKCE, one-hour access tokens and rotating refresh tokens (MCP guide). It is held to the workspace you chose: it can create and manage computers, screens and files there and read the workspace, and it never receives your computers' passwords. To disconnect a client, open Settings → Security in Orgo and click Disconnect next to it under Connected AI apps. Its next request fails, and the client asks you to sign in again.

A client that lets you set request headers can also send an API key to the same URL, as Authorization: Bearer <your key>. An API key reaches every tool its scope allows (MCP guide). Over the hosted URL, orgo_upload_file is refused, because it reads a file from the machine running the server. It works only with the local server.

How do I install it locally with an API key?

Create a key under Settings → API keys in Orgo, then add the server. It needs Node.js 20.3 or later (orgo-mcp-server on npm).

# Claude Code
claude mcp add orgo -e ORGO_API_KEY=sk_live_... -- npx -y orgo-mcp-server
 
# Codex
codex mcp add orgo --env ORGO_API_KEY=sk_live_... -- npx -y orgo-mcp-server

Claude Desktop and Cursor take the same JSON block. As of October 2026, Claude Desktop reads ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows, and Cursor reads ~/.cursor/mcp.json, or .cursor/mcp.json in a project (MCP: connect local servers, Cursor MCP docs):

{
  "mcpServers": {
    "orgo": {
      "command": "npx",
      "args": ["-y", "orgo-mcp-server"],
      "env": { "ORGO_API_KEY": "sk_live_..." }
    }
  }
}

A project's .cursor/mcp.json is often committed to the repository. In that file, write "ORGO_API_KEY": "${env:ORGO_API_KEY}" and keep the key in your environment.

VS Code reads .vscode/mcp.json, where the top-level key is servers rather than mcpServers. This block asks for the key when the server starts (MCP guide):

{
  "inputs": [
    { "type": "promptString", "id": "orgo-key", "description": "Orgo API key", "password": true }
  ],
  "servers": {
    "orgo": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "orgo-mcp-server"],
      "env": { "ORGO_API_KEY": "${input:orgo-key}" }
    }
  }
}

If you already ran orgo login with the Orgo CLI, leave out the env block. The server reads the same ~/.orgo/credentials.json, so the key never sits in a config file. Restart the client after you edit its config, since most clients read MCP config only at startup. On Windows, if the client reports spawn npx ENOENT, set command to cmd and start args with "/c", "npx" (MCP guide).

What can an agent do through Orgo MCP?

The MCP guide suggests these prompts to start with:

• spin up an orgo computer and screenshot the orgo.ai homepage
• open chrome on my orgo computer, search for the weather, screenshot the result
• list my orgo computers and stop the ones that are idle
• run the test suite on a clean orgo computer and show me what failed

For the first one, the model calls orgo_create_computer, then orgo_execute_bash, then orgo_get_screenshot (MCP guide).

To let the model see the screen, mind how screenshots come back. Over a hosted sign-in, orgo_get_screenshot always returns an image the model can see. With an API key, the tool returns a link to the stored capture by default, and the model cannot open that link. Tell the model to pass response_format: "base64", and the screenshot comes back as an image (MCP guide).

The local server has four settings worth knowing. None is required:

These belong to the local server. Over the hosted URL, pin a computer with the X-Orgo-Default-Computer-Id header, or add ?computer_id= to the URL for clients that only take a URL.

Orgo MCP, API, or CLI: which should I use?

They are complementary. All three reach the same computers, and the Python and TypeScript SDKs wrap the same HTTP API (Introduction).

When should I use Orgo MCP instead of a computer use loop?

Use Orgo MCP when your agent already lives in an MCP client such as Claude Code, Codex, Cursor, VS Code or Claude Desktop, and you want to ask for computer work in plain language without writing a tool router.

Run your own computer use loop with a model provider's tool when the agent lives in your application code and you need control over the model, retries and product behavior. See Run Claude computer use on Orgo, OpenAI computer use on Orgo and Gemini computer use on Orgo. Same computers, different calling convention.

To have Orgo run the loop for you, send a task to the OpenAI-compatible chat completions endpoint with a computer_id (Orgo computer use API).

You can mix them: MCP for hands-on work in your editor, the API for production workers.

How do I keep it safe?

• Leave your client's tool confirmation on. orgo_execute_bash and orgo_execute_python run any command on a live computer (MCP guide). A click or a keypress can also submit a form or delete something in any app on it.
• Use a workspace-scoped key. An sk_live_ key reaches every workspace by default. A key made for one workspace under Settings → API keys reaches only that workspace's computers.
• Narrow the local server. ORGO_READ_ONLY=true registers only the read tools, and ORGO_TAGS limits the tool groups.
• Prefer a hosted sign-in for one workspace. It is held to the workspace you picked and cannot reach templates, other workspaces, billing or computer passwords.
• Watch local file uploads. With the local server, orgo_upload_file reads a file from your own machine and uploads it to Orgo, so confirm those calls too.

How do I evaluate Orgo MCP for an agent stack?

1. Confirm you need a full computer, with a desktop, browser and shell, not only a browser tab. If the task stays inside a website, read Orgo vs Browserbase and Cloud desktops vs. headless browsers.
2. Connect the hosted URL or the local server as described above.
3. Start with the first prompt from the list: create a computer, run a command, take a screenshot.
4. Tighten access: tool confirmation on, a workspace-scoped key, and ORGO_READ_ONLY or ORGO_TAGS on the local server where they fit.
5. Check sizes and prices on Orgo pricing. If each of your users gets a computer, read How to scale AI agent computers to thousands.

FAQ

Is Orgo MCP an AI agent?

No. Orgo provides the computers and the MCP tools to drive them, and the model in your client decides what to do. Bring Claude, GPT, Gemini, Hermes Agent, OpenClaw, or your own loop (Introduction).

Do I need an API key for Orgo MCP?

Not for the hosted sign-in at https://www.orgo.ai/mcp. The local server, npx -y orgo-mcp-server, uses ORGO_API_KEY or the credentials saved by orgo login.

How many tools does Orgo MCP have?

With an API key, one per API operation. The list is generated from the published spec, so it grows with the API, and the local server prints the count when it starts. A hosted sign-in sees fewer: the tools for computers, screens and files in its workspace, plus reads of that workspace. Templates and workspace changes need an API key.

Can the model see the computer's screen?

Yes. A hosted sign-in always gets screenshots as images. With an API key, pass response_format: "base64". By default the tool returns a link to the stored capture (MCP guide).

Is Orgo MCP safer than giving an agent a shell on my laptop?

Commands run on a cloud computer that you create and can stop or delete, not on your own machine, so a bad command lands there instead of on your laptop. It does not replace tool confirmation: orgo_execute_bash and orgo_execute_python run any command on that computer, and the local server's orgo_upload_file reads files from your machine.

Does MCP replace a browser tool like Browserbase?

No. MCP is how your client reaches Orgo. Whether a task needs a browser or a full computer is a separate choice. See Orgo vs Browserbase.

Which MCP clients work with Orgo?

The Orgo docs show setups for Claude Code, Codex, Claude Desktop, Cursor and VS Code (MCP guide). A client that can start a local server takes the same four fields: npx, -y orgo-mcp-server, your key, and the stdio transport. Clients that support remote MCP servers can use the hosted URL.

Related

• What is computer use?
• Computer use models compared
• How do you pause, fork, or clone an AI agent's computer?
• Orgo templates and triggers