WinScript
State-aware Windows automation MCP server for AI agents, enabling app control, UI interaction, COM Office, and more via 59 tools with fallback layers and audit logging.
README
<div align="center">
██╗ ██╗██╗███╗ ██╗███████╗ ██████╗██████╗ ██╗██████╗ ████████╗
██║ ██║██║████╗ ██║██╔════╝██╔════╝██╔══██╗██║██╔══██╗╚══██╔══╝
██║ █╗ ██║██║██╔██╗ ██║███████╗██║ ██████╔╝██║██████╔╝ ██║
██║███╗██║██║██║╚██╗██║╚════██║██║ ██╔══██╗██║██╔═══╝ ██║
╚███╔███╔╝██║██║ ╚████║███████║╚██████╗██║ ██║██║██║ ██║
╚══╝╚══╝ ╚═╝╚═╝ ╚═══╝╚══════╝ ╚═════╝╚═╝ ╚═╝╚═╝╚═╝ ╚═╝
AppleScript for Windows. Built for AI agents.
Windows 10/11 · Python 3.10+ · MCP Protocol
</div>
macOS has AppleScript.
Windows had nothing clean for AI agents.
Until now.
WinScript is a state-aware, replayable, audited Windows automation server for AI agents. It wraps 4 fragmented Windows automation primitives — UI Automation, COM, Win32, and OCR — into a single MCP server that any agent can call.
Not a wrapper. Not a toy. Infrastructure.
Quick Start — Get WinScript Running in Claude Desktop
Option 1: Claude Desktop Extension (Easiest — Coming Soon)
Once approved in Claude's Extensions directory:
- Open Claude Desktop
- Go to Settings → Extensions
- Search for "WinScript"
- Click Install
- 59 tools appear — done!
Until then: Use Option 2 or 3 below.
Option 2: One-Click Installer
Step 1: Download this repo and double-click install.bat:
git clone https://github.com/RavaniRoshan/winscript-mcp.git
cd winscript-mcp
Then double-click install.bat (or run python install.py)
Step 2: Restart Claude Desktop
Step 3: WinScript appears in Claude's Extensions panel with 59 tools.
Option 3: PyPI (One Command)
pip install winscript
winscript
Then configure Claude Desktop manually (see below).
Option 4: Docker (Isolated)
docker run -v %USERPROFILE%/.winscript:~/.winscript ghcr.io/roshandamm/winscript-mcp:latest
Option 5: Direct from Source
git clone https://github.com/RavaniRoshan/winscript-mcp.git
cd winscript-mcp
pip install -r requirements.txt
python winscript-server.py
All options start an MCP server. The Claude Desktop Extension (Option 1) will be the easiest once approved.
The difference
Every other Windows automation tool gives you actions.
WinScript gives you actions + state.
# What others give you:
click("Submit")
→ "Clicked Submit"
# What WinScript gives you:
click("Submit")
→ "Clicked 'Submit' via uia_name [confidence 1.0] |
Active window: 'Form' → 'Confirmation' |
New windows: ['Success Dialog'] |
Duration: 312ms"
You don't just know what you did. You know what changed.
Detailed Installation
Option 1: Install from PyPI
pip install winscript
Then run: winscript or python -m winscript.server
Option 2: Run with Docker
# Pull and run
docker run -d --name winscript \
-v %USERPROFILE%/.winscript:~/.winscript \
ghcr.io/roshandamm/winscript-mcp:latest
# Or build locally
docker build -t winscript:latest .
docker run -d --name winscript -v %USERPROFILE%/.winscript:~/.winscript winscript:latest
Option 3: Run from Source (No Install)
git clone https://github.com/roshandamm/winscript-mcp.git
cd winscript-mcp
pip install -r requirements.txt
python winscript-server.py
Optional: OCR Fallback (Layer 4)
For better element detection in broken UI trees:
# Install Tesseract: https://github.com/tesseract-ocr/tesseract
pip install pytesseract
How WinScript Appears in Claude Desktop
After running the installer and restarting Claude Desktop, WinScript appears in Claude's Extensions panel just like Desktop Commander:
┌─────────────────────────────────────────────┐
│ WinScript │
│ AppleScript for Windows. Built for AI │
│ agents. Control any Windows app from Claude │
│ Enabled │
│ │
│ Developed by Roshan Ravani │
│ │
│ Tools 59 │
│ open_app │
│ close_app │
│ click │
│ type_text │
│ excel_read_cell │
│ outlook_send_email │
│ take_screenshot │
│ +53 more │
│ │
│ Requirements │
│ All requirements met │
│ │
│ Details │
│ Version 0.1.0 │
│ License MIT │
│ Author Roshan Ravani │
└─────────────────────────────────────────────┘
Claude can now:
- Open and control any Windows app
- Click buttons and type in UIs
- Read/write Excel files via COM
- Send Outlook emails
- Take screenshots
- Manage files and folders
- Record and replay workflows
- And 50+ more actions
All through natural language — no human interaction needed.
Wire into Claude Desktop
The easy way: Run install.bat — it configures everything for you.
The manual way: Edit %APPDATA%\Claude\claude_desktop_config.json:
{
"mcpServers": {
"winscript": {
"command": "python",
"args": ["-m", "winscript.server"]
}
}
}
Restart Claude Desktop. 59 tools appear automatically.
Five things that make WinScript different
1. Five-layer selector fallback chain
Other tools fail when the UI tree is bad (Electron apps, UWP, legacy Win32).
WinScript tries 5 strategies before giving up.
Layer 1 → UIA by element name (fast, exact)
Layer 2 → UIA by automation_id (for apps that label controls)
Layer 3 → UIA fuzzy role match (partial name, control type)
Layer 4 → OCR scan + bounding box (when UI tree is broken)
Layer 5 → Raw coordinates (click("x=412,y=308"))
Every tool call tells you which layer succeeded:
"Clicked 'Login' in 'Slack' [via ocr, confidence 0.91]"
2. State diffing after every action
Before you act, WinScript snapshots the desktop.
After you act, it snapshots again and diffs.
# The agent knows what actually happened:
open_app("excel")
→ "Opened Excel | Active window: '' → 'Book1 - Excel' |
New windows: ['Microsoft Excel - Book1'] | Duration: 2140ms"
type_text("Notepad", "hello")
→ "Typed 5 chars | No window change detected | Duration: 89ms"
No more "did it work?" loops.
3. Workflow recorder and replay
Record any successful multi-step sequence. Replay it on demand.
No human-written macros. No brittle scripts.
# Record:
workflow_record_start("daily_report", "Opens report and emails it")
open_latest_file("C:/reports", "xlsx")
read_active_document()
send_email_with_content("team@co.com", "Daily Report", "clipboard")
workflow_record_stop()
→ "Workflow 'daily_report' saved: 3 steps"
# Replay any time:
workflow_replay("daily_report")
→ "Step 1 ✓ open_latest_file → Opened q1_2026.xlsx [2100ms]
Step 2 ✓ read_active_document → [clipboard content] [340ms]
Step 3 ✓ send_email_with_content → Email sent [890ms]"
# Preview before running:
workflow_replay("daily_report", dry_run=True)
4. Semantic intent layer
Five high-level intents so agents don't have to think in clicks.
open_latest_file("C:/reports", "xlsx") # Find + open newest xlsx
send_email_with_content("a@b.com", "Re", "clipboard") # Clipboard → email
find_in_folder("C:/docs", "invoice", "pdf") # Find matching files
read_active_document() # Select-all copy current doc
summarize_screen() # Screenshot → agent vision
5. Full audit log + local memory
Every action, input, output, state delta, selector layer, and failure logged to ~/.winscript/audit.db.
get_audit_log(10)
→ "[14:23:01] ✓ open_app({'name':'notepad'}) → Opened notepad [2100ms]
[14:23:03] ✓ type_text({'text':'hello'}) → Typed 5 chars [89ms]
[14:23:11] ✗ click({'element':'Submit'}) → ERROR: No element found [412ms]"
get_failure_report()
→ "click: 3/12 failures (25%) | avg 380ms
open_app: 0/8 failures (0%) | avg 2100ms"
And memory persists across sessions:
what_files_have_i_opened(5, "xlsx")
→ "C:/reports/q1_2026.xlsx — opened 4x | last: 14:23 08/04"
what_did_i_do(5)
→ "[14:23] open_app → Opened notepad
[14:22] excel_read_cell → 47230.5
[14:21] outlook_send_email → Email sent to team@co.com"
All 59 tools
<details> <summary><b>App Control (4)</b></summary>
| Tool | What it does |
|---|---|
open_app(name) |
Open any app by name or alias |
close_app(title_hint) |
Close by partial window title |
focus_app(title_hint) |
Bring to foreground |
get_running_apps() |
List all open windows + PIDs |
</details>
<details> <summary><b>UI Interaction (5)</b></summary>
| Tool | What it does |
|---|---|
click(app_title, element_name) |
Click element — 5-layer fallback |
type_text(app_title, text) |
Type into focused element |
read_text(app_title, element_name) |
Read text from element |
press_key(key, app_title) |
Keyboard shortcuts |
get_ui_tree(app_title, depth) |
Discover all UI elements |
</details>
<details> <summary><b>COM Office (5)</b></summary>
| Tool | What it does |
|---|---|
excel_read_cell(filepath, sheet, cell) |
Read one cell |
excel_write_cell(filepath, sheet, cell, value) |
Write one cell + save |
excel_read_range(filepath, sheet, start, end) |
Read range as CSV |
outlook_send_email(to, subject, body) |
Send email |
outlook_read_inbox(count) |
Read N recent emails |
</details>
<details> <summary><b>File System (7)</b></summary>
read_file_text · write_file_text · list_dir · move_file · copy_file · delete_file · file_exists
</details>
<details> <summary><b>Screen + Clipboard (4)</b></summary>
| Tool | What it does |
|---|---|
take_screenshot(region) |
Base64 PNG — agent sees your screen |
get_active_window() |
Current focused window title |
get_clipboard() |
Read clipboard |
set_clipboard(text) |
Write clipboard |
</details>
<details> <summary><b>App Adapters (15)</b></summary>
Typed semantic APIs for specific apps. No more clicking blind.
# Excel
excel_open(filepath) · excel_save() · excel_close(save)
# Chrome
chrome_open(url) · chrome_navigate(url) · chrome_get_url()
chrome_get_title() · chrome_new_tab() · chrome_close_tab()
chrome_find_on_page(text)
# Notepad
notepad_open(filepath) · notepad_type(text)
notepad_save() · notepad_close(save)
# Explorer
explorer_open(path) · explorer_navigate(path)
# Outlook
outlook_open()
</details>
<details> <summary><b>Workflow Recorder + Replay (6)</b></summary>
workflow_record_start(name, description)
workflow_record_stop()
workflow_record_discard()
workflow_replay(name, dry_run)
workflow_list()
workflow_delete(name)
</details>
<details> <summary><b>Semantic Intents (5)</b></summary>
open_latest_file(folder, extension)
send_email_with_content(to, subject, content_source)
find_in_folder(folder, search_term, extension)
read_active_document()
summarize_screen()
</details>
<details> <summary><b>Audit + Memory + State (10)</b></summary>
# Audit
get_audit_log(limit, tool_filter)
get_failure_report()
# Memory
what_windows_have_i_seen(limit)
what_files_have_i_opened(limit, extension)
what_did_i_do(limit)
# State
get_state_snapshot()
# Modes
set_execution_mode(mode) # "safe" | "standard"
get_execution_mode()
</details>
App aliases
open_app("notepad") # notepad.exe
open_app("chrome") # chrome.exe
open_app("firefox") # firefox.exe
open_app("edge") # msedge.exe
open_app("excel") # EXCEL.EXE
open_app("word") # WINWORD.EXE
open_app("outlook") # OUTLOOK.EXE
open_app("explorer") # explorer.exe
open_app("terminal") # wt.exe
open_app("vscode") # Code.exe
open_app("cursor") # Cursor.exe
Error handling
Tools return "ERROR: ..." strings to the agent on failure — never crash your agent.
After 5 consecutive identical failures on the same tool + args:
WinScriptMaxRetriesError is raised. Hard stop. Change your args and try again.
get_failure_report()
# See which tools are failing and why before they hit the limit
Execution modes
set_execution_mode("safe")
# Read-only: screenshots, reads, audits only
# Blocks: write, delete, click, type, send email, open apps
set_execution_mode("standard")
# Full access (default)
Where recordings live
~/.winscript/
├── audit.db # every action ever taken
├── memory.db # windows, files, action history
└── workflows/
├── daily_report.json
└── your_workflow.json
Auto-purge: audit logs older than 30 days are deleted on startup.
Limitations (honest)
- Windows only. By design. This is not a bug.
- Elevated (admin) apps cannot be automated from a non-admin process.
- UWP + Electron apps have broken accessibility trees. WinScript falls back to OCR then coordinates — but complex UIs still sometimes fail.
- Requires Tesseract for OCR fallback (Layer 4). Without it, WinScript skips to Layer 5.
- COM automation (Excel, Outlook) requires those apps installed and licensed.
Built on
| Layer | Library |
|---|---|
| MCP server | FastMCP |
| UI automation | pywinauto + uiautomation |
| COM automation | pywin32 |
| OCR fallback | pytesseract + Tesseract |
| Screenshots | mss + Pillow |
| State + memory | SQLite |
Docker
# Pull from GitHub Container Registry
docker run -d --name winscript -v ${HOME}/.winscript:/root/.winscript ghcr.io/roshandamm/winscript-mcp:latest
# Or build from source
git clone https://github.com/RavaniRoshan/winscript-mcp.git
cd winscript-mcp
docker build -t winscript:latest .
docker run -d --name winscript -v winscript_data:/root/.winscript winscript:latest
# Using Docker Compose
docker-compose up -d
Build Claude Desktop Extension
# Windows: Double-click build-extension.bat
# Linux/Mac: ./build-extension.sh
This creates winscript.mcpb - a one-click installable extension for Claude Desktop.
Test locally: Double-click the .mcpb file → Claude Desktop opens → review → install
Developer Commands
# Build package
pip install build && python -m build
# Install locally
pip install -e .
# Run tests
pip install pytest && pytest tests/ -v
# Connect to Cursor
# Edit .cursor/mcp.json:
{"mcpServers": {"winscript": {"command": "python", "args": ["-m", "winscript.server"]}}}
Examples
Screenshot + Vision Fallback
For legacy apps with broken UI trees:
open_app("Legacy Inventory")take_screenshot()→ pass to Claude Visioncoordinate_click(x, y)to click visually located element
Excel to Email Workflow
excel_read_range("report.xlsx", "Sheet1", "A1:D10")outlook_send_email("team@company.com", "Daily Report", body)
File Triage
list_dir("C:/Downloads")move_file("file.pdf", "C:/Documents/")
Troubleshooting
"Server disconnected" errors:
- Ensure PYTHONPATH is set correctly in Claude Desktop config
- Check Python path is valid
"Module not found: winscript":
- Run
python winscript-server.pyfrom the project root - Or use
pip install -e .to install the package
COM automation fails (Excel/Outlook):
- Ensure Microsoft Office is installed and licensed
- Run as administrator for first-time COM registration
UI elements not found:
- Try increasing timeout:
wait_for_window("App Name", timeout_seconds=30) - Enable OCR fallback: install Tesseract and
pip install pytesseract
License
MIT
<div align="center">
WinScript — 59 tools. State-aware. Replayable. Audited. Memory-backed.
Built by Roshan Ravani
</div>
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.