MCP Hub
Back to servers

Linux Desktop MCP Server

Enables AI assistants to control Linux desktop applications via AT-SPI2 accessibility interfaces, allowing for semantic element targeting, natural language UI search, and cross-platform input simulation.

Tools
8
Updated
Jan 8, 2026
Validated
Jan 9, 2026

Linux Desktop MCP Server

PyPI version License: MIT Python 3.10+

Built with Claude Code - This entire MCP server was developed using Claude Code, Anthropic's AI-powered coding assistant. We're proud to showcase what's possible with AI-assisted development!

An MCP server that provides Chrome-extension-level semantic element targeting for native Linux desktop applications using AT-SPI2 (Assistive Technology Service Provider Interface).

Features

  • Semantic Element References: Just like Chrome extension's ref_1, ref_2 system
  • Role Detection: Identifies buttons, text fields, links, menus, etc.
  • State Detection: Tracks focused, enabled, checked, editable states
  • Natural Language Search: Find elements by description ("save button", "search field")
  • Cross-Platform Input: Works on X11, Wayland, and XWayland
  • GTK/Qt/Electron Support: Works with any application that exposes accessibility

Installation

System Dependencies

# Ubuntu/Debian
sudo apt install python3-pyatspi gir1.2-atspi-2.0 at-spi2-core

# For X11 input simulation
sudo apt install xdotool

# For Wayland input simulation (recommended)
# Install ydotool from source or your package manager
# Then start the daemon:
sudo ydotoold &

Python Package

# From PyPI
pip install linux-desktop-mcp

# Or from source
git clone https://github.com/yourusername/linux-desktop-mcp.git
cd linux-desktop-mcp
pip install -e .

Enable Accessibility

Ensure accessibility is enabled in your desktop environment:

  • GNOME: Settings → Accessibility → Enable accessibility features
  • KDE: System Settings → Accessibility
  • Most modern desktops have this enabled by default

Configuration

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "linux-desktop": {
      "command": "linux-desktop-mcp"
    }
  }
}

Or if installed from source:

{
  "mcpServers": {
    "linux-desktop": {
      "command": "python",
      "args": ["-m", "linux_desktop_mcp"]
    }
  }
}

Available Tools

desktop_snapshot

Capture the accessibility tree with semantic element references.

Parameters:
  app_name: str (optional) - Filter to specific application
  max_depth: int (default: 15) - Tree traversal depth

Returns:
  Tree of elements with ref_ids:
  - ref_1: [application] Firefox
    - ref_2: [frame] "GitHub - Mozilla Firefox"
      - ref_3: [button] "Back" (clickable)
      - ref_4: [entry] "Search or enter address" (editable, focused)

desktop_find

Find elements by natural language query.

Parameters:
  query: str - "save button", "search field", "menu containing File"
  app_name: str (optional)

Returns:
  Matching elements with refs, states, and actions

desktop_click

Click an element by reference or coordinates.

Parameters:
  ref: str - Element reference (e.g., "ref_5")
  element: str - Human description for logging
  coordinate: [x, y] - Fallback if no ref
  button: left|right|middle
  click_type: single|double
  modifiers: [ctrl, shift, alt, super]

desktop_type

Type text into an element.

Parameters:
  text: str - Text to type
  ref: str - Element to focus first (optional)
  element: str - Human description
  clear_first: bool - Ctrl+A, Delete before typing
  submit: bool - Press Enter after

desktop_key

Press keyboard keys/shortcuts.

Parameters:
  key: str - Key name (Return, Tab, Escape, a, etc.)
  modifiers: [ctrl, shift, alt, super]

desktop_capabilities

Check available automation capabilities.

Example Usage

Example 1: Navigating to a Website in Firefox

User: "Open GitHub in Firefox"

Claude uses:
1. desktop_snapshot(app_name="Firefox")
   → Returns UI tree with elements like:
     - ref_5: [entry] "Search or enter address" (editable, focused)
     - ref_12: [button] "Go" (clickable)

2. desktop_click(ref="ref_5", element="URL bar")
   → Clicks to focus the address bar

3. desktop_type(text="https://github.com", ref="ref_5", clear_first=True, submit=True)
   → Types the URL and presses Enter

Result: Firefox navigates to GitHub

Example 2: Saving a File in LibreOffice

User: "Save this document as 'report.odt'"

Claude uses:
1. desktop_key(key="s", modifiers=["ctrl"])
   → Opens the Save dialog

2. desktop_snapshot(app_name="LibreOffice")
   → Returns dialog elements including:
     - ref_8: [entry] "File name:" (editable)
     - ref_15: [button] "Save" (clickable)

3. desktop_type(text="report.odt", ref="ref_8", clear_first=True)
   → Types the filename

4. desktop_click(ref="ref_15", element="Save button")
   → Clicks Save

Result: Document saved as report.odt

Example 3: Searching in a Code Editor

User: "Search for 'TODO' comments in VS Code"

Claude uses:
1. desktop_find(query="search", app_name="Code")
   → Finds search-related elements

2. desktop_key(key="f", modifiers=["ctrl", "shift"])
   → Opens global search panel

3. desktop_snapshot(app_name="Code")
   → Returns search panel elements:
     - ref_22: [entry] "Search" (editable, focused)
     - ref_25: [checkbox] "Match Case"

4. desktop_type(text="TODO", ref="ref_22", submit=True)
   → Types search query and executes search

Result: VS Code shows all TODO comments across the project

Example 4: Window Targeting for Multi-Window Automation

User: "Help me copy data from the spreadsheet to the email"

Claude uses:
1. desktop_context(list_available=True)
   → Lists all available windows

2. desktop_target_window(app_name="LibreOffice Calc", color="green")
   → Targets spreadsheet with green border

3. desktop_target_window(app_name="Thunderbird", color="blue")
   → Targets email client with blue border

4. desktop_snapshot()
   → Only shows elements from targeted windows (reduced context)

5. [Proceeds with copy/paste operations between windows]

Result: Claude can efficiently work across multiple applications

Platform Support

FeatureX11WaylandXWayland
AT-SPI discoveryFullFullFull
Click by refFullFullFull
Type textFullFullFull
ydotool inputFullFullFull
xdotool inputFullNoYes

Troubleshooting

"AT-SPI2 not available"

sudo apt install python3-pyatspi gir1.2-atspi-2.0 at-spi2-core

"AT-SPI2 registry not running"

Ensure accessibility is enabled in your desktop settings. You may need to log out and back in.

"No input backend available" (Wayland)

# Install and start ydotool daemon
sudo ydotoold &

Elements not showing up

Some applications may not expose accessibility information. Modern GTK3/4, Qt5/6, and Electron apps generally work well.

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     MCP Protocol Layer                       │
│              (JSON-RPC over stdio, tool defs)                │
└─────────────────────────────────────────────────────────────┘
                              │
┌─────────────────────────────────────────────────────────────┐
│                  Reference Manager                           │
│          (ref_1, ref_2 mapping, lifecycle, GC)              │
└─────────────────────────────────────────────────────────────┘
                              │
          ┌───────────────────┴───────────────────┐
          │                                       │
┌─────────────────────┐             ┌─────────────────────────┐
│   AT-SPI2 Backend   │             │   Input Backends        │
│     (pyatspi)       │             │ (ydotool/xdotool/wtype) │
└─────────────────────┘             └─────────────────────────┘

Contributing

This project was created with Claude Code and we warmly welcome contributions! Whether you want to:

  • Report bugs or request features
  • Submit pull requests
  • Fork and build your own version
  • Improve documentation

We're very open to help and collaboration. See CONTRIBUTING.md for guidelines.

Privacy Policy

Linux Desktop MCP is a local desktop automation tool that:

  • Runs entirely on your local machine - No data is transmitted to external servers
  • Does not collect any personal data - No analytics, telemetry, or usage tracking
  • Does not store credentials - All authentication and authorization is handled by your local system
  • Accesses only what you explicitly target - The accessibility tree is read only for windows/applications you interact with
  • No network connectivity required - The MCP server operates completely offline

The only data accessed is the accessibility tree information exposed by your desktop applications (UI element names, roles, and states), which is used solely for local automation and is not persisted or transmitted anywhere.

Contact: For privacy-related questions, open an issue on GitHub.

License

MIT - See LICENSE for details.

Reviews

No reviews yet

Sign in to write a review