# realhands — Real Desktop Computer Use [Health: Active]

**Category:** 🖥️ OS Automation  
**Repository:** https://github.com/kanishka089/computer-use-mcp  
**GitHub Stars:** 1  
**Views:** 2  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/realhands-real-desktop-computer-use

## Description
Lets Claude drive your REAL mouse, keyboard & screen — works in your own logged-in apps.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "realhands-real-desktop-computer-use": {
    "command": "uvx",
    "args": ["realhands@latest"],
    "env": {
      "COMPUTER_USE_MAX_DIM": "",
      "COMPUTER_USE_IMAGE_FORMAT": "",
      "CHANGE_THRESHOLD": "",
      "MAX_SKIPS": "",
      "COMPUTER_USE_IDLE_STOP": ""
    }
  }
}
```

**Requires environment variables:** `COMPUTER_USE_MAX_DIM`, `COMPUTER_USE_IMAGE_FORMAT`, `CHANGE_THRESHOLD`, `MAX_SKIPS`, `COMPUTER_USE_IDLE_STOP` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What realhands — Real Desktop Computer Use does

realhands — Real Desktop Computer Use connects Claude Desktop or Claude Code to the physical desktop session running the MCP server. It can read the screen, move the real cursor, click, drag, scroll, type, press keys, hold keys, activate a window, wait for an interface to settle, and report cursor or monitor information. These actions work with the user's existing applications and logged-in browser sessions rather than a separate browser profile.

The server is intended for direct computer control. It does not provide an abstract browser API or a separate automation sandbox. Actions affect the machine attached to the session, so it should only be used where autonomous input to that desktop is acceptable.

## How it works

Claude drives the interaction by calling the single `computer` tool. A typical task starts with `screenshot`; Claude interprets the returned image, chooses coordinates, and sends the next action. After a non-screenshot action, the server waits briefly for the interface and normally returns another screenshot. A `steps` list can combine multiple actions and return one final image, reducing repeated image transmission. A failed step stops the batch and reports the failure.

Screenshots are resized to a configurable maximum dimension and aligned to 28-pixel vision patches. Coordinates are mapped back to real screen pixels, with clamping to the selected monitor. On Windows, DPI awareness keeps screenshot coordinates aligned with pointer coordinates under display scaling. The optional `monitor` argument selects the primary display, another display, or the full virtual desktop; monitor enumeration is available through the tool.

## Setup and configuration

The repository includes `install.py`, which creates a virtual environment, installs dependencies, creates configuration, and can register the server with Claude Desktop. The implementation uses Python 3.10 or 3.11, FastMCP over stdio, and desktop input and capture libraries including PyAutoGUI, MSS, Pillow, pynput, keyboard, pyperclip, python-dotenv, and Windows-specific window-management packages.

Configuration can be supplied through an optional `.env` file. `COMPUTER_USE_MAX_DIM` controls screenshot sizing, while `COMPUTER_USE_IMAGE_FORMAT` selects the image format. `CHANGE_THRESHOLD` and `MAX_SKIPS` control suppression of screenshots when the screen has changed little. `COMPUTER_USE_IDLE_STOP` can enable automatic stand-down after inactivity; it is disabled by default.

## Tools and capabilities

The `computer` tool supports:

- Screen capture, cursor-position reporting, and monitor enumeration
- Mouse movement, clicks, double and triple clicks, dragging, button state, and scrolling
- Text entry, key presses, key chords, and timed key holds
- Window activation by title substring
- Delayed waits and explicit task shutdown with `stop`
- Batched action sequences through `steps`

realhands — Real Desktop Computer Use also has three documented safety mechanisms: moving the pointer to the fail-safe corner, a configurable panic hotkey, and an always-on-top STOP overlay. Safety controls arm on the first action and stand down when `stop` is called.

## Limitations and notes

This server is fully autonomous and acts on the real machine without asking for approval before each input. Keep the panic controls available while it runs. The fail-safe corner and STOP behavior terminate or abort activity, while the panic hotkey hard-kills the process.

Screenshots consume vision context and are returned after actions unless suppressed for a batch. Smaller images reduce cost but may reduce grounding detail. The same monitor must be used for the screenshot and the coordinates derived from it. On multi-monitor systems, screen origins can be negative when a display is positioned left of or above the primary display.

The documented client integrations are Claude Desktop and Claude Code.

_Full upstream README: https://allmcps.com/mcp/realhands-real-desktop-computer-use/readme_