Windows Desktop Automation v2

Windows Desktop Automation

End-to-end Windows desktop automation for AI coding agents. Drive legacy applications, terminal emulators, Citrix windows, and GUI-only tools with the same rigor you apply to REST APIs.

What it does

This skill gives your AI agent a structured workflow for automating any Windows desktop application — even those with no CLI, no API, and no documented automation surface. It combines four proven techniques into one resilient stack:

• UI Automation (pywinauto + UIA) for native Win32, WPF, WinForms, and Qt apps
• Hotkey/Macro layer for keyboard-shortcut-driven workflows
• Image recognition (OpenCV + Tesseract OCR) for Citrix, RDP, custom-drawn, and remote-desktop UIs
• Resilience patterns — retry, recovery, evidence capture, and human-in-the-loop fallback

It enforces a "sniff → classify → select → execute → verify" cycle before any click or keypress, dramatically reducing the silent-failure problem that kills most RPA scripts.

When to use it

• The target app has no API, no CLI, and no automation hooks
• You need to drive Citrix, RDP, terminal emulators, or VB6/MFC legacy apps
• Selenium/Playwright can't help (the UI is not in a browser)
• You need screen-scraping + form-filling with audit trail
• A human currently does the task and you want to hand it off to an agent

Why it's better than ad-hoc prompting

Most LLMs will try to pyautogui.click(540, 320) once and call it done. This skill is different:

• It inspects the window tree first (Inspect.exe, uiautomation, pywinauto) before any action
• It classifies the target control (Edit, Button, List, Tree, etc.) and picks the right interaction mode
• It uses multiple locator strategies in parallel (UIA, name, class, image hash, OCR) and falls back gracefully
• It captures screenshots before/after every step for postmortem
• It never trusts a click — verifies state change with a follow-up read

Architecture

┌──────────────────────────────────────────────────────────┐
│                    Agent (Claude/Cursor)                 │
│  - Translates intent into action plan                    │
│  - Calls helper scripts via Bash                         │
└───────────────┬──────────────────────────────────────────┘
                │
                ▼
┌──────────────────────────────────────────────────────────┐
│             skills/windows-desktop-automation/           │
│  scripts/                                                │
│    ├── sniff_window.py     # UIA + window tree dump      │
│    ├── classify_target.py  # Control type + locator plan │
│    ├── act.py              # Click/type/scroll + verify  │
│    ├── ocr_click.py        # Image+OCR fallback          │
│    ├── hotkey.py           # Global hotkey macros         │
│    └── evidence.py         # Screenshot + log capture    │
│  references/                                              │
│    ├── locator-strategies.md                              │
│    ├── citrix-rdp-gotchas.md                              │
│    └── uia-cheatsheet.md                                 │
└──────────────────────────────────────────────────────────┘

Quick start

# 1. Install
pip install pywinauto opencv-python pytesseract pillow psutil

# 2. Sniff a window to discover its controls
python scripts/sniff_window.py --title "Order Entry"

# 3. Drive an action
python scripts/act.py --window "Order Entry" \
  --action type --control "Customer ID" --value "ACME-001"

# 4. If UIA fails, fall back to image+OCR
python scripts/ocr_click.py --window "Order Entry" \
  --template templates/submit_button.png

Supported scenarios

Installation

Claude Code / OpenClaw / Codex CLI

npx agensi install windows-desktop-automation

Manual

Download windows-desktop-automation.zip and unzip to ~/.claude/skills/ (or your agent's skills directory).

Example usage

"Open SAP GUI, navigate to transaction VA01, create a sales order for customer ACME-001, and dump the resulting order number to a JSON file."

The skill will:

1. Launch SAP GUI and wait for the main window
2. Sniff the tree, locate the transaction code field
3. Type "VA01", press Enter, wait for the new screen
4. Re-sniff, locate the Order Type field
5. Drive the multi-step form with locator+image fallback
6. Capture the resulting order number from the status bar
7. Write JSON + screenshots to ./evidence/

Common failure modes (and how this skill handles them)

• Window not found → polls every 500ms up to 30s, then OCR fallback
• Control not enabled → waits + retries, then captures evidence
• Modal dialog appears → auto-detects via window-tree diff, dismisses or escalates
• Stale element reference → re-sniffs and re-locates
• OS DPI scaling mismatch → normalizes via GetDpiForWindow
• Locked workstation → checks GetForegroundWindow, refuses to run

Pricing

Single-purchase, lifetime access. Includes:

• All 6 helper scripts
• 3 reference docs (locator strategies, Citrix gotchas, UIA cheatsheet)
• 5 example workflows
• Future updates for the same major version

Support

GitHub-style issues aren't open for marketplace items. For help, contact via the creator's page — typically a 48-hour response window.

Compatibility

Works with any agent that supports the SKILL.md standard and can execute Python: Claude Code, OpenClaw, Codex CLI, Cursor, Gemini CLI, Cline, Windsurf, Aider. Tested on Windows 10/11, Server 2019+.

Tags

windows, automation, desktop, ui-automation, pywinauto, ocr, rpa, devops, citrix, rdp, legacy