MCP server
by ab498
Automate desktop tasks with Computer Control: mouse, keyboard, screenshots, OCR & window management. Power Automate Desk
Provides desktop automation capabilities including mouse control, keyboard input, screen capture, and OCR text recognition. Works directly with your computer's GUI without external dependencies.
Computer Control is a community-built MCP server published by ab498 that provides AI assistants with tools and capabilities via the Model Context Protocol. Automate desktop tasks with Computer Control: mouse, keyboard, screenshots, OCR & window management. Power Automate Desk It is categorized under developer tools.
You can install Computer Control in your AI client of choice. Use the install panel on this page to get one-click setup for Cursor, Claude Desktop, VS Code, and other MCP-compatible clients. This server runs locally on your machine via the stdio transport.
MIT
Computer Control is released under the MIT license. This is a permissive open-source license, meaning you can freely use, modify, and distribute the software.
Add new capabilities to Claude beyond text generation
Example
Access external data sources, execute code, interact with tools and services
Transform Claude from chatbot to action-taking agent
Provide Claude with access to relevant context and data
Example
Load project documentation, access knowledge bases, query databases
Get more accurate, context-aware responses
Automate multi-step workflows combining AI and external tools
Example
Research → Summarize → Create document → Send notification
Complete complex tasks end-to-end without manual steps
Share your MCP server with the developer community
Computer Control is among the better-indexed MCP projects we tried; the explainx.ai summary tracks the official description.
Computer Control reduced integration guesswork — categories and install configs on the listing matched the upstream repo.
We wired Computer Control into a staging workspace; the listing’s GitHub and npm pointers saved time versus hunting across READMEs.
We evaluated Computer Control against two servers with overlapping tools; this profile had the clearer scope statement.
Computer Control has been reliable for tool-calling workflows; the MCP profile page is a good permalink for internal docs.
According to our notes, Computer Control benefits from clear Model Context Protocol framing — fewer ambiguous “AI plugin” claims.
I recommend Computer Control for teams standardizing on MCP; the explainx.ai page compares cleanly with sibling servers.
Strong directory entry: Computer Control surfaces stars and publisher context so we could sanity-check maintenance before adopting.
According to our notes, Computer Control benefits from clear Model Context Protocol framing — fewer ambiguous “AI plugin” claims.
Computer Control has been reliable for tool-calling workflows; the MCP profile page is a good permalink for internal docs.
showing 1-10 of 36
uvx)Note: Running uvx computer-control-mcp@latest for the first time will download python dependencies (around 70MB) which may take some time. Recommended to run this in a terminal before using it as MCP. Subsequent runs will be instant.
{
"mcpServers": {
"computer-control-mcp": {
"command": "uvx",
"args": ["computer-control-mcp@latest"]
}
}
}
OR install globally with pip:
pip install computer-control-mcp
Then run the server with:
computer-control-mcp # instead of uvx computer-control-mcp, so you can use the latest version, also you can `uv cache clean` to clear the cache and `uvx` again to use latest version.
Traditional screenshot methods like GDI/PrintWindow fail to capture GPU-accelerated windows, resulting in black screens. This impacts games, media players, Electron apps, browsers with GPU acceleration, streaming software, and CAD tools. Use WGC through take_screenshot tool's flag or ENV variable
By default, screenshots are saved to the OS downloads directory. You can customize this by setting the COMPUTER_CONTROL_MCP_SCREENSHOT_DIR environment variable:
{
"mcpServers": {
"computer-control-mcp": {
"command": "uvx",
"args": ["computer-control-mcp@latest"],
"env": {
"COMPUTER_CONTROL_MCP_SCREENSHOT_DIR": "C:\Users\YourName\Pictures\Screenshots"
}
}
}
}
Or set it system-wide:
# Windows (PowerShell)
$env:COMPUTER_CONTROL_MCP_SCREENSHOT_DIR = "C:\Users\YourName\Pictures\Screenshots"
# macOS/Linux
export COMPUTER_CONTROL_MCP_SCREENSHOT_DIR="/home/yourname/Pictures/Screenshots"
If the specified directory doesn't exist, the server will fall back to the default downloads directory.
You can configure the system to automatically use Windows Graphics Capture (WGC) for specific windows by setting the COMPUTER_CONTROL_MCP_WGC_PATTERNS environment variable. This variable should contain comma-separated patterns that match window titles:
{
"mcpServers": {
"computer-control-mcp": {
"command": "uvx",
"args": ["computer-control-mcp@latest"],
"env": {
"COMPUTER_CONTROL_MCP_WGC_PATTERNS": "obs, discord, game, steam"
}
}
}
}
Or set it system-wide:
# Windows (PowerShell)
$env:COMPUTER_CONTROL_MCP_WGC_PATTERNS = "obs, discord, game, steam"
# macOS/Linux
export COMPUTER_CONTROL_MCP_WGC_PATTERNS="obs, discord, game, steam"
When this variable is set, any window whose title contains any of the specified patterns will automatically use WGC for screenshot capture, eliminating black screens for GPU-accelerated applications.
click_screen(x: int, y: int): Click at specified screen coordinatesmove_mouse(x: int, y: int): Move mouse cursor to specified coordinatesdrag_mouse(from_x: int, from_y: int, to_x: int, to_y: int, duration: float = 0.5): Drag mouse from one position to anothermouse_down(button: str = "left"): Hold down a mouse button ('left', 'right', 'middle')mouse_up(button: str = "left"): Release a mouse button ('left', 'right', 'middle')type_text(text: str): Type the specified text at current cursor positionpress_key(key: str): Press a specified keyboard keykey_down(key: str): Hold down a specific keyboard key until releasedkey_up(key: str): Release a specific keyboard keypress_keys(keys: Union[str, List[Union[str, List[str]]]]): Press keyboard keys (supports single keys, sequences, and combinations)take_screenshot(title_pattern: str = None, use_regex: bool = False, threshold: int = 60, scale_percent_for_ocr: int = None, save_to_downloads: bool = False, use_wgc: bool = False): Capture screen or windowtake_screenshot_with_ocr(title_pattern: str = None, use_regex: bool = False, threshold: int = 10, scale_percent_for_ocr: int = None, save_to_downloads: bool = False): Extract adn return text with coordinates using OCR from screen or windowget_screen_size(): Get current screen resolutionlist_windows(): List all open windowsactivate_window(title_pattern: str, use_regex: bool = False, threshold: int = 60): Bring specified window to foregroundwait_milliseconds(milliseconds: int): Wait for a specified number of milliseconds# Clone the repository
git clone https://github.com/AB498/computer-control-mcp.git
cd computer-control-mcp
# Build/Run:
# 1. Install in development mode | Meaning that your edits to source code will be reflected in the installed package.
pip install -e .
# Then Start server | This is equivalent to `uvx computer-control-mcp@latest` just the local code is used
computer-control-mcp
# -- OR --
# 2. Build after `pip install hatch` | This needs version increment in orer to reflect code changes
hatch build
# Windows
$latest = Get-ChildItem .\dist\*.whl | Sort-Object LastWriteTime -Descending | Select-Object -First 1
pip install $latest.FullName --upgrade
# Non-windows
pip install dist/*.whl --upgrade
# Run
computer-control-mcp
python -m pytest
See the API Reference for detailed information about the available functions and classes.
MIT
Prerequisites
Time Estimate
15-60 minutes depending on server complexity
Steps
Troubleshooting
✓ Do
✗ Don't
💡 Pro Tips
Architecture
Model Context Protocol standardizes how AI hosts (Claude, Cursor) communicate with external tools and data sources through server implementations.
Protocols
Compatibility
✓ Use when
Use when you need Claude to access external data, execute actions, or integrate with tools. Best for extending AI capabilities beyond conversation.
✗ Avoid when
Avoid when native integrations exist (use official APIs directly), for real-time critical systems, or when security/compliance requires zero external dependencies.