Agentfenster

What the agentfenster MCP server does

The agentfenster MCP server exposes real Windows GUI control through a local MCP process. Applications run on a separate Win32 desktop inside the current Windows session, so the agent can work without moving the user’s foreground window, mouse, or keyboard. The desktop can still be watched in the agentfenster app.

The server is intended for Windows software that an agent must operate through its visible interface. Because it runs inside the user’s session, applications can use the session’s existing logins, cookies, and VPN connection. It is not an open-source implementation: the MCP component is distributed with the commercial, closed-source agentfenster app.

How it works

The server uses Windows UI Automation element trees rather than screenshots as its primary source of interface information. Window content is returned in compact form, with a maximum of 200 elements per window. Agents can first discover available windows, inspect a target, and then select controls by identifier or name.

For higher-level work, desktop_task runs a complete GUI task within the server and returns a concise result describing success or the failed step. Its internal retries and intermediate actions do not enter the model context. Lower-level tools expose the individual operations: opening an application, reading a desktop window, clicking, entering text, pressing keys, and scrolling.

After an action, the server checks the relevant window state. A click or keystroke that produces no observable change is reported as such rather than being treated as successful. replay provides recent run history, including plans, actions, observed changes, and a link to visual replay in the app.

Setup and configuration

The agentfenster MCP server requires Windows 10 or 11 and Python 3.11 or newer supplied through the app. Install agentfenster from its product site; the MCP server is included with that installation. The product begins with a three-day trial and is commercial software.

Register the local stdio process with Claude Code using:

claude mcp add agentfenster -- agentfenster-mcp

Other clients can use the same server if they support launching a local stdio MCP process. The agent desktop starts without applications, so use open_app before trying to interact with a program. The server itself does not list required environment variables. The model endpoint used by the internal desktop_task click loop must be configured, but the supplied material does not specify its configuration method or variable names.

Tools and capabilities

The agentfenster MCP server includes these 11 tools:

• list_windows discovers controllable windows and reports titles, classes, minimized state, and capture availability.
• desktop_task performs a complete task in one window and reports the outcome.
• inspect_window returns a filtered, capped UI element tree for planning.
• replay lists recent runs or shows a run’s detailed action log.
• open_app launches a program on the agent desktop.
• read_desktop lists interactive elements in an agent-desktop window.
• click selects an element by ID or name and verifies the result.
• type_text enters text into an element and checks for a change.
• press_key sends keys or combinations, optionally focusing an element first.
• scroll moves a window or pane and reports whether its content moved.
• agent_windows lists open agent-desktop windows, with the frontmost first.

Limitations and notes

The agentfenster MCP server is local and uses stdio transport. Risky operations such as deleting, sending, paying, or installing are blocked by default for MCP calls and reported instead of waiting for an unavailable confirmation response.

The server reports no telemetry. According to the supplied documentation, its only network traffic is the model endpoint configured for the desktop_task click loop. The repository contains documentation and issue tracking rather than the product source code, and using the application is governed by its commercial terms.