Agentfenster MCP server

What it does

agentfenster MCP server drives Windows GUI applications on a hidden second desktop, controlled by Claude Code or other MCP clients. The server runs on Windows 10/11 and uses UI Automation element trees instead of screenshots, with verification of every action against actual window state.

• Automate repetitive tasks in Windows applications while keeping your screen and input free for other work.
• Run full GUI workflows in a single call via desktop_task, which handles retries internally without consuming context.
• Inspect and interact with specific windows by reading their element tree and executing verified clicks, text input, and key presses.

Tools this server exposes

Tool	Does
list_windows	Lists available windows with title, class, minimized state and capture capability.
desktop_task	Runs a complete GUI task and returns a summary of what happened and whether it succeeded.
inspect_window	Returns the element tree of a window as a compact list, capped at 200 elements.
replay	Lists recent runs or returns a step-by-step log of one run with visual replay link.
open_app	Launches a program on the agent desktop and reports its first titled window.
read_desktop	Reads the compact element tree of a window, one line per interactive element.
click	Clicks an element by id or name and verifies whether anything changed.
type_text	Types text into an element and verifies whether anything changed.
press_key	Presses a key or key combination, optionally focusing an element first, with verification.
scroll	Scrolls a window or pane and reports whether the content actually moved.
agent_windows	Lists windows currently open on the agent desktop, frontmost first.

Before you enable it

An MCP server runs with whatever access you give it. Start with a development account rather than production, and check which tools the server exposes before you let an agent call them unattended.