🖥️ PC-Control MCP Server

An MCP tool server that gives Claude Desktop control of a Windows machine — shell execution, system stats, audio and power, screenshots, window and process management.

How it works

The server is the tool side of an agent loop. It does not decide when to act; it publishes a set of callable tools and executes them when the client asks.

Claude Desktop  ──MCP/stdio──▶  FastMCP server (main.py)
                                      │
                                      ├── tools/shell.py         execute_command
                                      ├── tools/system.py        get_system_status
                                      ├── tools/audio.py         manage_volume
                                      ├── tools/power.py         manage_power
                                      ├── tools/app_launcher.py  launch_application
                                      ├── tools/processes.py     get_process_list
                                      ├── tools/screen.py        take_screenshot
                                      └── tools/window.py        list_windows, manage_window

Each tool is registered with an @mcp.tool() decorator over a plain typed function:

@mcp.tool()
def system_manage_volume(action: str, level: int = None) -> str:
    """
    Control system audio volume.
    Args:
        action: 'set', 'get', 'mute', 'unmute'
        level: 0-100 (required for 'set')
    """
    return manage_volume(action, level)

The signature and docstring are the tool schema — that is how the client discovers what exists, what arguments it takes, and when to call it. Which means the docstring is an interface contract, not a comment: a vague one produces a tool the model calls at the wrong time.

Two design choices worth naming:

• Transport separated from behaviour. main.py holds only registration; every tool body lives in tools/. So verify_tools.py can exercise the tools directly, without an LLM in the loop — when something misbehaves, that separates "the tool is broken" from "the model called it wrong."

• Destructive actions are explicit. Power management takes an action enum that includes abort, so a shutdown started by mistake can be cancelled.

web-frontend-mcp/ is a second, separate server covering browser-side tooling.

⚠️ This server executes arbitrary shell commands on the host by design. Run it only against a client you control.

Verification

verify_tools.py and verify_new_tools.py call each tool directly and print the result. These are smoke checks, not a test suite — no assertions, no fixtures, no CI.

Related MCP server: SysControl MCP

Quick Install

1. Clone or download this folder

2. Install dependencies:

pip install -r requirements.txt

3. Configure Claude Desktop:

Edit: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "pc-control": {
      "command": "python",
      "args": ["C:/path/to/pc-control-mcp/main.py"]
    }
  }
}

4. Restart Claude Desktop

Usage

Ask Claude:

• "Check my system status"

• "Execute command: dir"

• "Set volume to 50%"

• "Open Chrome"

• "Lock my computer"

Requirements

• Python 3.8+

• Windows 10/11

• Claude Desktop App

Tools

system_execute_command

Execute any shell command

system_get_status

Get CPU, Memory, Disk, Battery status

system_manage_volume

Control audio volume (set, get, mute, unmute)

system_manage_power

Power actions (shutdown, restart, sleep, lock)

system_launch_app

Launch applications by name

system_get_processes

List running processes sorted by CPU or Memory

screen_take_screenshot

Capture screenshot of primary monitor

window_list

List specific details of all visible windows

window_manage

Manage windows (focus, minimize, maximize, close)

License

MIT

Created with ❤️ for Windows automation