Desktop Control Skill

Control the desktop GUI using semantic element targeting. Find and click UI elements by name instead of coordinates.

Key Features:

• AT-SPI - Primary method using accessibility tree (knows element roles, states, actions)
• OCR Fallback - Tesseract-based text finding when AT-SPI can't find the element
• Wait Utilities - Poll for elements to appear with exponential backoff
• Click Verification - Optional pre-click screenshot verification

Prerequisites

Install dependencies:

bash install.sh

Or manually:

# System packages
sudo apt-get install -y xdotool scrot imagemagick \
    at-spi2-core libatk-adaptor python3-gi gir1.2-atspi-2.0 \
    tesseract-ocr tesseract-ocr-eng python3-pip

# Python packages
pip3 install -r requirements.txt

For headless Xvfb sessions:

export GTK_MODULES=gail:atk-bridge
export QT_LINUX_ACCESSIBILITY_ALWAYS_ON=1
/usr/lib/at-spi2-core/at-spi-bus-launcher &

Commands

All commands use DISPLAY=:10.0 by default. Override with --display flag.

find-element

Find UI element via AT-SPI with OCR fallback.

python3 scripts/desktop.py find-element --name "Confirm" [--role button] [--app Firefox]

Returns:

{
  "element": {
    "name": "Confirm",
    "bounds": { "x": 400, "y": 300, "width": 100, "height": 30 },
    "center": { "x": 450, "y": 315 },
    "role": "push button",
    "source": "atspi",
    "visible": true,
    "enabled": true,
    "clickable": true
  }
}

find-text

Find text on screen via OCR only.

python3 scripts/desktop.py find-text "I have an existing wallet" [--exact] [--all]

Returns:

{
  "match": {
    "name": "I have an existing wallet",
    "bounds": { "x": 200, "y": 400, "width": 180, "height": 20 },
    "center": { "x": 290, "y": 410 },
    "source": "ocr",
    "confidence": 95.2
  }
}

click-element

Click element by name/role (finds element first, then clicks at center).

python3 scripts/desktop.py click-element --name "Next" [--role button] [--verify]

click-element requires at least one selector: --name or --role. When --verify is used, OCR must be available and text must be provided (typically via --name).

Returns:

{
  "clicked": {
    "element": { "name": "Next", "..." },
    "x": 450,
    "y": 315,
    "button": "left",
    "double": false
  }
}

wait-for

Wait for element or text to appear (with timeout and exponential backoff).

python3 scripts/desktop.py wait-for --name "Success" --timeout 30
python3 scripts/desktop.py wait-for --text "Transaction complete" --timeout 60
python3 scripts/desktop.py wait-for --name "Loading" --gone --timeout 30

Use either --text or element selectors (--name, --role, --app) for a single call, not both. For element waits (with or without --gone), provide at least one of --name or --role.

Returns:

{ "found": { "name": "Success", "..." } }
// or
{ "gone": true, "name": "Loading" }
// or
{ "error": "Element not found within 30s", "timeout": true }

list-elements

List all interactive elements (buttons, inputs, links, etc.)

python3 scripts/desktop.py list-elements [--app Firefox] [--role button]

Returns:

{
  "elements": [
    { "name": "Sign In", "role": "push button", "..." },
    { "name": "Email", "role": "entry", "..." }
  ],
  "count": 2
}

status

Check AT-SPI and OCR availability.

python3 scripts/desktop.py status

Returns:

{
  "atspi": {
    "available": true,
    "applications": ["Firefox", "gnome-calculator"]
  },
  "ocr": { "available": true },
  "display": ":10.0"
}

Original Commands

These coordinate-based commands are still available:

Example Workflows

MetaMask Transaction (Semantic)

# Wait for and click Confirm button
python3 scripts/desktop.py wait-for --name "Confirm" --role button --timeout 30
python3 scripts/desktop.py click-element --name "Confirm" --role button

# Wait for success message
python3 scripts/desktop.py wait-for --text "Transaction submitted" --timeout 60

Phantom Wallet Import (Semantic)

# Click "I have an existing wallet"
python3 scripts/desktop.py click-element --name "I already have a wallet"

# Wait for seed phrase input
python3 scripts/desktop.py wait-for --role entry --timeout 10

# Type seed phrase
python3 scripts/desktop.py type "word1 word2 word3..."

# Click Import
python3 scripts/desktop.py click-element --name "Import"

Hybrid Approach (Semantic + Coordinates)

# Use semantic for known buttons
python3 scripts/desktop.py click-element --name "Settings"

# Fall back to coordinates for unlabeled icons
python3 scripts/desktop.py screenshot --output /tmp/screen.png
# (analyze screenshot to get coordinates)
python3 scripts/desktop.py click 850 120

Tips

1. Prefer semantic commands - click-element and wait-for are more robust than coordinates
2. Check status first - Run status to verify AT-SPI and OCR are available
3. Use --role for precision - Distinguish between buttons and text with same name
4. Fall back to OCR - If AT-SPI doesn't expose an element, find-text uses OCR
5. Wait instead of sleep - wait-for is more reliable than fixed delays
6. Use --verify for critical clicks - Adds OCR verification before clicking