@superpms/jarvis-remote-hand-mcp

An MCP service that lets AI agents remotely control multiple computers under your account to execute system commands. Typical use case: troubleshooting and fixing system issues (like network problems) on multiple machines at once.

This service does not log in by itself — it borrows the existing login state of the "Remote Control Assistant" desktop app: agent → this service → desktop app → MQTT → target device.

Prerequisites

1. A machine running the "Remote Control Assistant" desktop app, logged in

2. Target devices also running the desktop app, logged into the same account, and online

3. Node.js >= 20.19

Related MCP server: mcp-windows-server

Configuration

Add the following to your MCP client configuration:

{
  "mcpServers": {
    "jarvis-remote-hand": {
      "command": "npx",
      "args": ["-y", "@superpms/jarvis-remote-hand-mcp"],
      "env": {
        "ARH_BRIDGE_URL": "http://127.0.0.1:47890",
        "ARH_BRIDGE_TOKEN": "在桌面端「MCP 服务配置」页复制"
      }
    }
  }
}

Environment variables

About the access token

The bridge interface is equivalent to a shell proxy for the logged-in device and listens on the LAN for cross-machine use, so the server enforces token validation. The token is stored in the desktop app's application data directory and persists across restarts; to revoke an old token, click "Reset" on the "MCP Service Configuration" page — after that, all agent configurations must be updated with the new value.

health is the only interface that does not require a token — it exists precisely so you can get an answer to "is the bridge actually running" before the token is even paired, so it only returns a fixed {"ok":true} with no other information.

Cross-machine access also requires the target machine's firewall to allow port 47890.

Tools

The tool parameter of send_command is chosen based on the target device's platform: bash (macOS/Linux), cmd or pws (Windows).

Usage flow

1. list_devices to get the device_uuid of an online device

2. send_command to send, get the mid

3. Poll query_mid until state is "Completed", then read result

state goes through Sending → Queued → Running → Completed in order. The controlled end executes commands one by one in the order received; only one runs at a time.

Command execution has no timeout — anywhere from a few minutes to over ten minutes is normal (e.g. network resets, batch installs). When polling shows "Running", wait a few seconds and continue; do not judge it as failed or re-send.

If a device is offline or not under the current account, send_command returns an error explaining the reason — it will not return a mid that never produces a result.

Security notes

Requests forwarded by this service can execute arbitrary system commands on your logged-in devices. Use only on trusted networks; do not expose port 47890 to the public internet, and do not post the access token in public places.

License

MIT