MCP server for Windows desktop automation and control, providing tools for mouse/keyboard input, window management, screenshot capture, web scraping, PowerShell execution, and UI element interaction
Windows-MCP.Net provides 24 tools with consistent naming conventions (verb-first: Click, Drag, Type, etc.) and functional descriptions. However, the server has significant gaps in schema completeness, parameter documentation, and error handling guidance. Most tools lack documented output schemas, and several parameter descriptions are minimal. The codebase shows good use of logging and dependency injection, but does not fully align with production agentic tool patterns around idempotency, error recovery, and composition.
Click on UI elements at specific coordinates
Copy text to clipboard or retrieve current clipboard content
Copy a file from source to destination
Drag and drop operation from source coordinates to destination coordinates
Execute PowerShell commands and return the output with status code
Find UI element by automation ID
Find UI element by class name
Find UI element by text content
No documented output schemas for any tool. LLMs cannot plan multi-step sequences or extract required downstream IDs (e.g., what does GetWindowInfoAsync return? window bounds? center coordinates?). Schema scores are severely penalized across all tools.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 50 | <=2025-11-25 | v2 |
| 2026-03-09 | D | 50 | - | v1 |
Capture comprehensive desktop state including applications, UI elements, and content
Get properties of UI element at specified coordinates
Get window position and size information for a specific application window
Press individual keyboard keys
Launch an application from the Windows Start Menu by name
Move mouse cursor to specific coordinates without clicking
Open a URL in the default browser
Resize a specific application window to specific size or move to specific location
Fetch and convert webpage content to markdown format
Scroll at specific coordinates or current mouse position
Execute keyboard shortcuts using key combinations
Switch to a specific application window and bring it to foreground
Take a screenshot and save it to the temp directory
Type text into input fields, text areas, or focused elements
Pause execution for specified duration in seconds
Wait for UI element to appear with specified selector
ExecuteCommandAsync (PowerShell) and CopyFileAsync are destructive tools with no error recovery guidance. Descriptions do not clarify whether operations are idempotent, whether failed commands should be retried, or what compensation steps exist. No permission gates or audit trails visible.
Many parameter descriptions are minimal (<50 chars) and lack constraint information. Examples: KeyAsync's 'key' parameter lists special keys in quotes but provides no enum or validation rules; ExecuteCommandAsync's 'command' has no guidance on what commands are safe/allowed; OpenBrowserAsync's 'url' parameter default is null with no explanation of what happens when omitted.
No tool annotations detected (readOnlyHint, destructiveHint, idempotentHint). MCP servers should declare which tools are read-only, destructive, or idempotent so agents can plan safely. Currently, ClickAsync, DragAsync, TypeAsync, and others are marked as WRITE or DESTRUCTIVE in the assessment but have no formal annotations in code.
TakeScreenshotAsync has no input parameters documented, and its output (file path? base64 image? metadata?) is not specified. This blocks agents from using the screenshot in downstream tools.
UI element search tools (FindElementByTextAsync, FindElementByClassNameAsync, FindElementByAutomationIdAsync) do not document what happens when multiple matches exist, whether they return coordinates for ClickAsync, or how the agent should distinguish between similar elements.
No error recovery guidance. If GetWindowInfoAsync returns 'Failed to get window info', the code suggests trying the app name in the default language, but this error message is hardcoded in the tool, not returned from the underlying service. LLMs cannot parse this and act on it correctly.
ClipboardAsync's 'text' parameter is marked 'required for copy mode' but is optional (default=null). LLMs may omit it in copy mode and receive a null-reference error rather than a clear validation message.