winremote-mcp 0.4.23

Windows Remote MCP Server - control Windows desktops via MCP protocol

WinRemote MCP — Run MCP Servers Remotely on Windows

The ultimate Windows MCP server for remote desktop control and automation. Control any Windows machine through the Model Context Protocol — perfect for AI agents, Claude Desktop. Transform your Windows desktop into a powerful, remotely-accessible automation endpoint.

Run on the Windows machine you want to control. Built with FastMCP and the Model Context Protocol.

Quickstart (30 seconds)

# Install from PyPI
pip install winremote-mcp

# Start the Windows MCP server
winremote-mcp

That's it! Your Windows MCP server is now running on http://127.0.0.1:8090 and ready to accept commands from MCP clients like Claude Desktop.

🤖 Agent Integrations

winremote-mcp works with MCP-compatible AI agents and clients. Client-specific setup guides live in this repo:

• Hermes — add winremote as a native streamable HTTP or stdio MCP server.
• OpenClaw — use winremote as OpenClaw's Windows control layer.
• Claude Desktop / Claude Code — configure local stdio or remote streamable HTTP.
• Cursor — add winremote to .cursor/mcp.json.

Hermes Setup

Run winremote-mcp on the Windows machine, then add it to Hermes as an MCP server:

pip install winremote-mcp
winremote-mcp --host 0.0.0.0 --port 8090 --auth-key "your-secret-key"

mcp_servers:
  winremote:
    type: streamable-http
    url: http://<windows-ip>:8090/mcp
    headers:
      Authorization: Bearer your-secret-key

See the full Hermes integration guide for local stdio setup, verification prompts, and available capabilities.

🤖 OpenClaw Integration

winremote-mcp is the official Windows control layer for OpenClaw. Together they give your AI agent full remote control over any Windows machine — screenshots, PowerShell, file transfer, GUI automation, and more.

---

The Easiest Way: Just Tell OpenClaw

You don't need to configure anything manually. Just tell your OpenClaw agent:

"Install winremote-mcp on my Windows machine at 192.168.1.100 and connect it to yourself. Python is installed at C:\Python311\python.exe."

OpenClaw will SSH into the Windows machine, install the package, start the server, and wire up the MCP connection — all automatically.

---

Manual Setup (Step by Step)

Step 1 — Install on Windows

pip install winremote-mcp

Step 2 — Start the server

Local-only quick start (default, safest):

winremote-mcp

Remote access requires authentication:

Starting with v0.4.20, winremote-mcp refuses to bind HTTP transport to a non-loopback address without authentication. Use an API key for LAN or remote access:

winremote-mcp --host 0.0.0.0 --port 8090 --auth-key YOUR_SECRET_KEY

For lab-only legacy behavior, you can explicitly acknowledge the risk with --allow-insecure-remote; do not use this on shared, routed, or internet-exposed networks.

Auto-start on boot:

winremote-mcp install

Step 3 — Connect OpenClaw

Add to your openclaw.json:

{
  "plugins": {
    "entries": {
      "winremote": {
        "type": "mcp",
        "url": "http://192.168.1.100:8090/mcp",
        "headers": {
          "Authorization": "Bearer YOUR_SECRET_KEY"
        }
      }
    }
  }
}

Or tell your OpenClaw agent:

"Add winremote MCP at http://192.168.1.100:8090/mcp with auth key YOUR_SECRET_KEY."

Step 4 — What your agent can do

Once connected, your AI agent has full Windows control:

Capability	Example
🖥️ Screenshots	Capture the full desktop or a specific window
⚡ Shell execution	Run PowerShell, CMD, or batch scripts
📁 File transfer	Upload/download files between Linux and Windows
🖱️ GUI automation	Click, type, drag — control any Windows app
🔧 System info	Process list, services, event logs, registry
📷 OCR	Extract text from any screen region
🎬 Screen recording	Record desktop activity as GIF

---

Secure Remote Access (HTTPS)

For access over the internet or untrusted networks, enable HTTPS:

Step 1 — Generate a certificate:

# Self-signed (LAN/homelab)
openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -days 365 -nodes

# Trusted cert (no browser warnings) — requires mkcert installed
mkcert -install && mkcert 192.168.1.100

Step 2 — Start with TLS:

winremote-mcp --host 0.0.0.0 --port 8090 ^
  --auth-key YOUR_SECRET_KEY ^
  --ssl-certfile cert.pem ^
  --ssl-keyfile key.pem

OpenClaw config with HTTPS:

{
  "plugins": {
    "entries": {
      "winremote": {
        "type": "mcp",
        "url": "https://192.168.1.100:8090/mcp",
        "headers": {
          "Authorization": "Bearer YOUR_SECRET_KEY"
        }
      }
    }
  }
}

---

OAuth 2.0 (for Claude Desktop and other MCP clients)

Some MCP clients (like Claude Desktop) use OAuth instead of API keys. OAuth is a pre-provisioned confidential-client flow: configure both client ID and client secret on the server, then copy the same values into the client. Dynamic client registration is disabled, and redirect URIs must be loopback http(s) URIs.

winremote-mcp --host 0.0.0.0 --port 8090 ^
  --ssl-certfile cert.pem --ssl-keyfile key.pem ^
  --oauth-client-id my-client --oauth-client-secret my-secret

Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "winremote": {
      "type": "http",
      "url": "https://192.168.1.100:8090/mcp/",
      "oauth": {
        "clientId": "my-client",
        "clientSecret": "my-secret"
      }
    }
  }
}

---

winremote.toml — Full Config Reference

Place in your working directory or ~/.config/winremote/winremote.toml:

[server]
host         = "0.0.0.0"
port         = 8090
auth_key     = "your-secret-key"
# Optional break-glass flag for legacy trusted-LAN deployments only.
# Remote HTTP without auth is refused unless this is true.
allow_insecure_remote = false
ssl_certfile = "C:/certs/cert.pem"   # optional — enables HTTPS
ssl_keyfile  = "C:/certs/key.pem"    # optional — enables HTTPS

[security]
ip_allowlist        = ["192.168.1.0/24"]   # restrict to LAN only
oauth_client_id     = ""                    # optional OAuth client ID
oauth_client_secret = ""                    # optional OAuth secret

[tools]
exclude = ["ScreenRecord"]   # disable specific tools

---

Note: winremote-mcp is a standard MCP server and works with any MCP-compatible client — Claude Desktop, Cursor, OpenClaw, and others.

What's New in v0.4.23

🐛 FastMCP debug/uvicorn compatibility

• Fixed winremote-mcp --debug with FastMCP 3.2.4+, where run_http_async renamed the uvicorn options keyword from uvicorn_args to uvicorn_config.
• winremote-mcp now detects the installed FastMCP signature and passes DEBUG uvicorn logging through the supported keyword, while preserving compatibility with older FastMCP releases.
• Added regression coverage so the debug HTTP startup path stays compatible with both keyword names.

中文发布说明

• 修复 FastMCP 3.2.4+ 中 run_http_async 参数从 uvicorn_args 改为 uvicorn_config 后，winremote-mcp --debug 启动时报错的问题。
• 现在会自动检测当前 FastMCP 支持的参数名，并继续把 uvicorn 的 DEBUG 日志配置传递给 HTTP 传输层；旧版 FastMCP 仍保持兼容。
• 已添加回归测试，覆盖新旧 uvicorn 配置参数的兼容路径。

What's New in v0.4.22

🐛 Debug flag restored

• Added the documented --debug CLI flag so winremote-mcp --debug is accepted.
• --debug enables DEBUG logging for winremote and passes log_level=debug to uvicorn for HTTP transport.

🔒 Dependency security updates

• Added minimum constraints for idna>=3.15 and starlette>=1.0.1 to avoid known vulnerable versions.

📚 README release notes cleanup

• README now keeps only the latest two What's New sections.
• Older release notes remain available in the full CHANGELOG.

What's New in v0.4.21

📚 README release notes cleanup

• README keeps recent What's New sections focused and points older release notes to the full CHANGELOG.

For older release notes, see the full CHANGELOG.

What Problem It Solves

• Remote Windows Control: Control Windows desktops from anywhere through standardized MCP protocol
• AI Agent Integration: Enable Claude, GPT, and other AI agents to interact with Windows GUI applications
• Cross-Platform Automation: Bridge the gap between Linux/macOS development environments and Windows targets
• Headless Windows Management: Manage Windows servers and workstations without RDP or VNC overhead

Features

• Desktop Control — Screenshot capture (JPEG compressed, multi-monitor), click, type, scroll, keyboard shortcuts
• Window Management — Focus windows, minimize-all, launch/resize applications, multi-monitor support
• Remote Shell Access — PowerShell command execution with working directory support
• File Operations — Read, write, list, search files; binary transfer via base64 encoding
• System Administration — Windows Registry access, service management, scheduled tasks, process control
• Network Tools — Ping hosts, check TCP ports, monitor network connections
• Advanced Features — OCR text extraction, screen recording (GIF), annotated screenshots with UI element labels
• AI Vision Support — Works with Flutter, Electron, Qt and any UI via AI vision. See Vision Guide
• Security & Auth — Optional API key authentication, localhost-only binding by default

Installation

From PyPI (Recommended)

pip install winremote-mcp

From Source

git clone https://github.com/dddabtc/winremote-mcp.git
cd winremote-mcp
pip install .

With Optional Dependencies

# Install with OCR support (includes pytesseract)
pip install winremote-mcp[ocr]

# Install development dependencies
pip install winremote-mcp[test]

OCR Setup (Optional)

For text extraction from screenshots:

# 1. Install Tesseract OCR engine
winget install UB-Mannheim.TesseractOCR

# 2. Install with OCR dependencies
pip install winremote-mcp[ocr]

Usage

Basic Usage

Tier and tool controls

# Default: tier1 + tier2 enabled, tier3 disabled
winremote-mcp

# Enable destructive tier3 tools
winremote-mcp --enable-tier3

# Disable interactive tier2 (tier1 only)
winremote-mcp --disable-tier2

# Both together: tier1 + tier3 (tier2 disabled)
winremote-mcp --enable-tier3 --disable-tier2

# Backward-compatible: enable everything
winremote-mcp --enable-all

# Explicit tool list (highest precedence over tier flags)
winremote-mcp --tools Snapshot,Click,Type

# Remove specific tools from resolved set
winremote-mcp --enable-tier3 --exclude-tools Shell,FileWrite

Config file (winremote.toml)

Search order:

1. --config /path/to/winremote.toml
2. ./winremote.toml
3. ~/.config/winremote/winremote.toml

[server]
host = "127.0.0.1"
port = 8090
auth_key = ""
ssl_certfile = ""       # Path to SSL certificate for HTTPS
ssl_keyfile = ""        # Path to SSL private key for HTTPS

[security]
ip_allowlist = ["127.0.0.1", "192.168.1.0/24"]
enable_tier3 = false
disable_tier2 = false
oauth_client_id = ""    # Expected OAuth client ID (optional)
oauth_client_secret = "" # OAuth client secret for confidential clients

[tools]
enable = ["Snapshot", "Click", "Type"]
exclude = []

Precedence: CLI flags override config file values; config file values override defaults.

IP allowlist

# CLI
winremote-mcp --ip-allowlist 127.0.0.1,192.168.1.0/24

# Or via config [security].ip_allowlist

Supports both single IPs and CIDR ranges (IPv4/IPv6). Non-allowlisted clients receive HTTP 403 with a clear error.

HTTPS / TLS

To enable HTTPS, provide SSL certificate and key files:

winremote-mcp --ssl-certfile cert.pem --ssl-keyfile key.pem

Or in winremote.toml:

[server]
ssl_certfile = "/path/to/cert.pem"
ssl_keyfile  = "/path/to/key.pem"

Generate a self-signed certificate (for local/LAN use):

openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -days 365 -nodes

OAuth 2.0

WinRemote MCP includes a built-in OAuth 2.0 Authorization Server, compatible with Claude Desktop and other MCP clients that require OAuth.

Enable it with:

winremote-mcp --oauth-client-id my-client --oauth-client-secret my-secret

Or in winremote.toml:

[security]
oauth_client_id     = "my-client"
oauth_client_secret = "my-secret"

Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "winremote": {
      "type": "http",
      "url": "https://your-host:8080/mcp/",
      "oauth": {
        "clientId": "my-client",
        "clientSecret": "my-secret"
      }
    }
  }
}

The OAuth server implements:

• GET /.well-known/oauth-authorization-server — server metadata (RFC 8414)
• POST /oauth/register — dynamic client registration (RFC 7591)
• GET /oauth/authorize — Authorization Code + PKCE (RFC 7636)
• POST /oauth/token — token exchange

Health check

# Start MCP server (localhost only, no auth)
winremote-mcp

# Start with remote access and authentication
winremote-mcp --host 0.0.0.0 --port 8090 --auth-key "your-secret-key"

# Enable all tools including high-risk Tier 3 (Shell, FileWrite, etc.)
winremote-mcp --enable-all

# Start with hot reload for development
winremote-mcp --reload

MCP Client Configuration

For Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "winremote": {
      "command": "winremote-mcp",
      "args": ["--transport", "stdio"]
    }
  }
}

For HTTP MCP clients:

{
  "mcpServers": {
    "winremote": {
      "type": "streamable-http", 
      "url": "http://192.168.1.100:8090/mcp",
      "headers": {
        "Authorization": "Bearer your-secret-key"
      }
    }
  }
}

Auto-Start on Boot

# Create Windows scheduled task
winremote-mcp install

# Remove scheduled task  
winremote-mcp uninstall

Security

Tools are organized into three risk tiers. By default, only Tier 1-2 tools are enabled.

Tier	Risk	Default	Examples
Tier 1	Read-only	✅ Enabled	Snapshot, GetSystemInfo, FileList
Tier 2	Interactive	✅ Enabled	Click, Type, Shortcut, Scrape
Tier 3	Destructive / server-side effects	❌ Disabled	Shell, App, PlaySound, FileWrite

# Enable all tiers (use with caution)
winremote-mcp --enable-all

# Always use auth for remote access
winremote-mcp --host 0.0.0.0 --auth-key "your-secret-key"

See SECURITY.md for the full security guide.

Tools

Tool	Description
Desktop
Snapshot	Screenshot (JPEG, configurable quality/max_width) + window list + UI elements
AnnotatedSnapshot	Screenshot with numbered labels on interactive elements
OCR	Extract text from screen via OCR (pytesseract or Windows built-in)
ScreenRecord	Record screen activity as animated GIF
PlaySound	Play audio file on Windows host (.wav/.mp3/.ogg/.wma/.m4a, local path or URL)
Input
Click	Mouse click (left/right/middle, single/double/hover)
Type	Type text at coordinates
Scroll	Vertical/horizontal scroll
Move	Move mouse / drag
Shortcut	Keyboard shortcuts
Wait	Pause execution
Window Management
FocusWindow	Bring window to front (fuzzy title match)
MinimizeAll	Show desktop (Win+D)
App	Launch/switch/resize applications
System
Shell	Execute PowerShell commands (with optional cwd)
GetClipboard	Read clipboard
SetClipboard	Write clipboard
ListProcesses	Process list with CPU/memory
KillProcess	Kill process by PID or name
GetSystemInfo	System information
Notification	Windows toast notification
LockScreen	Lock workstation
ReconnectSession	Reconnect disconnected Windows desktop session to console
File System
FileRead	Read file content
FileWrite	Write file content
FileList	List directory contents
FileSearch	Search files by pattern
FileDownload	Download file as base64 (binary)
FileUpload	Upload file from base64 (binary)
Registry & Services
RegRead	Read Windows Registry value
RegWrite	Write Windows Registry value
ServiceList	List Windows services
ServiceStart	Start a Windows service
ServiceStop	Stop a Windows service
Scheduled Tasks
TaskList	List scheduled tasks
TaskCreate	Create a scheduled task
TaskDelete	Delete a scheduled task
Network
Scrape	Fetch URL content
Ping	Ping a host
PortCheck	Check if a TCP port is open
NetConnections	List network connections
EventLog	Read Windows Event Log entries

How It Works

graph LR
    A["MCP Client<br/>(Claude/AI)"] -->|commands| B["WinRemote MCP<br/>Server"]
    B -->|API calls| C["Windows APIs<br/>(Win32/WMI/PS)"]
    C -->|results| B
    B -->|responses| A

Transport Options:

• stdio: Direct process communication (ideal for Claude Desktop)
• HTTP: RESTful API with optional authentication (ideal for remote access)

Core Architecture:

1. Tool Layer: 40+ Windows automation tools (screenshot, click, type, etc.)
2. Task Manager: Concurrency control and task cancellation
3. Transport Layer: MCP protocol over stdio or HTTP
4. Security Layer: Optional Bearer token authentication

Working with Non-Standard UI Frameworks

AnnotatedSnapshot uses Win32 API to detect UI elements, which doesn't work with Flutter, Electron, Qt, or custom-drawn UIs. Three solutions:

Approach	Setup	GPU	Best For
Snapshot + Claude Vision	None	No	Most users — Claude sees the screenshot and clicks
UI-TARS Desktop	Medium	16 GB	Highest accuracy (94.2%), best Chinese UI support
OmniMCP	Medium	16 GB	Multi-LLM setups (LLM-agnostic)

Quick example — no extra tools needed:

You:    "Take a screenshot with Snapshot, find the Connect button, and click it."
Claude:  1. Calls Snapshot() → sees the Flutter app screenshot
         2. Vision identifies "Connect" button at (520, 340)
         3. Calls Click(x=520, y=340)

For the complete guide with setup instructions, architecture diagrams, and comparison benchmarks, see docs/vision-guide.md.

Troubleshooting / FAQ

Q: MCP server not starting?

A: Check Python version (requires 3.10+) and ensure no other service is using port 8090:

python --version
netstat -an | findstr :8090

Q: Can't connect from remote machine?

A: Use --host 0.0.0.0 to bind to all interfaces (default is localhost only):

winremote-mcp --host 0.0.0.0 --auth-key "secure-key"

Q: Screenshot tool returns empty/black images?

A: Windows may be locked or display turned off. Ensure:

• Windows is unlocked and display is active
• No screen saver is running
• For multi-monitor setups, specify monitor parameter

Q: OCR not working?

A: Install Tesseract OCR engine:

winget install UB-Mannheim.TesseractOCR
pip install winremote-mcp[ocr]

Q: Permission errors with registry/services?

A: Run with administrator privileges:

# Right-click Command Prompt → "Run as administrator"
winremote-mcp

Contributing

We welcome contributions! Please see our Contributing Guide for details.

Development Setup

git clone https://github.com/dddabtc/winremote-mcp.git
cd winremote-mcp
pip install -e ".[test]"
pytest  # Run tests

Acknowledgments

Inspired by Windows-MCP by CursorTouch. Thanks for the pioneering work on Windows desktop automation via MCP.

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Ready to automate Windows with AI? ⚡ Install winremote-mcp and connect your favorite AI agent to any Windows machine in under 30 seconds.