MCP server

Gives an agent most of the API as tools (not key minting and revocation), from a server we host or one your client runs locally.

Hosted

The hosted server is the quickest way in. Point a client that supports remote MCP at https://app.mandala.computer/mcp; it opens a browser to approve access, and you can revoke it any time in Settings → Credentials → Connected apps. The access token it gets is for that endpoint only and is not an API key. A client that cannot open a browser, such as a CI job, can send a Mandala API key as Authorization: Bearer com_… instead, if it lets you set that header; it is scoped exactly as that key is.

claude mcp add --scope user --transport http mandala 'https://app.mandala.computer/mcp'

{
  "mcpServers": {
    "mandala": {
      "type": "http",
      "url": "https://app.mandala.computer/mcp"
    }
  }
}

Run it locally

To run it yourself instead: your MCP client starts mandala-computer-mcp@0.16.0 locally over stdio with npx; Node.js 20.3 or newer. Run the commands in a POSIX shell on your local machine, or merge the JSON entry into a client configuration that supports mcpServers. Replace YOUR_COMPUTER_ID with an existing computer’s stable ID and YOUR_MANDALA_API_KEY locally with a key from Settings → Credentials → API keys. If a server named mandala already exists, update its configuration in your client.

claude mcp add mandala \
  -e 'MANDALA_API_KEY=YOUR_MANDALA_API_KEY' \
  -e 'MANDALA_BASE_URL=https://app.mandala.computer/api/v1' \
  -e 'MANDALA_COMPUTER_ID=YOUR_COMPUTER_ID' \
  -- npx -y mandala-computer-mcp@0.16.0

codex mcp add mandala \
  --env 'MANDALA_API_KEY=YOUR_MANDALA_API_KEY' \
  --env 'MANDALA_BASE_URL=https://app.mandala.computer/api/v1' \
  --env 'MANDALA_COMPUTER_ID=YOUR_COMPUTER_ID' \
  -- npx -y mandala-computer-mcp@0.16.0

{
  "mcpServers": {
    "mandala": {
      "command": "npx",
      "args": [
        "-y",
        "mandala-computer-mcp@0.16.0"
      ],
      "env": {
        "MANDALA_API_KEY": "YOUR_MANDALA_API_KEY",
        "MANDALA_BASE_URL": "https://app.mandala.computer/api/v1",
        "MANDALA_COMPUTER_ID": "YOUR_COMPUTER_ID"
      }
    }
  }
}

MANDALA_COMPUTER_ID selects a default target; tools can override it. It does not restrict the key’s authority. A key reaches its account with its holder’s current permissions, narrowed to a workspace if one was selected when creating it. Keep keys out of conversations and revoke them in Settings when no longer needed. Source, tool list and skill →

First task

After configuration, ask the agent to call screenshot with computer_id set to your computer’s stable ID, and describe the screen before acting. Start or resume the computer when ready to use a current screenshot or shell.

Each computer page has a Connect an agent action with these instructions, and the CLI shell, filled in for that computer and a Context option for sharing its ID, specifications and API reference. Context alone does not install tools or authenticate. Opening the panel or copying text creates no key and does not start the computer; a key is made only when you choose Create API key, which reaches every computer in the account, or Create workspace key for a computer in a workspace. If authentication fails, check the key and API base; replace an invalid, expired or revoked credential in Settings. For access denied, check the key’s account and workspace and your current role.