agentfenster

com.agentfensterPublished by agentfenster.comv0.1.0Updated Oct 27, 2026

Windows GUI automation on a hidden second desktop, driven from Claude Code over MCP.

VerifiedSTDIODesktop onlyProductivity & WorkflowBrowser Automation

Overview

AI-generated overview

Local Windows MCP server that lets an assistant drive real desktop applications on a hidden second desktop, with verified clicks, typing and key presses.

What you can do
  • List and inspect windows: enumerate drivable windows, read a compact UI Automation element tree, and list windows open on the shared agent desktop
  • Launch programs on the agent desktop and report their first titled window
  • Act on a window: click elements, type text, press key combinations and scroll, each verified against the resulting window state
  • Run a whole GUI task in one call, with the click loop executed inside the server and returned as a single summary paragraph
  • Review past runs: list recent runs or fetch a step-by-step log with a link to the visual replay
When to use it
Useful when an assistant should operate a Windows application that has no API, such as filling in a legacy tool or navigating a desktop app, while the user keeps their own screen, mouse and keyboard. Suited to Windows 10 or 11 users who want GUI automation with explicit success or failure reporting.
Requirements
Windows 10 or 11 only, with the commercial agentfenster app installed; the MCP server ships with the app. Local stdio server launched as agentfenster-mcp, needing Python 3.11 or newer as installed by the app. A model endpoint must be configured in the app for the desktop_task click loop. No environment variables or headers are declared.
Before you install
The app is commercial closed-source software; every plan starts with a 3-day trial and then requires payment. Risky actions such as delete, send, pay or install are denied by default on MCP calls, so some tasks will be reported as blocked rather than completed. The configured model endpoint is the only network traffic, so prompts and screen-derived element data go to that endpoint. Actions change real application state.

Installation

In SourceWeft

  1. Open agentfenster in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

agentfenster MCP server

The agentfenster MCP server lets Claude Code and other MCP clients drive real Windows applications on a hidden second desktop inside your own session, with the same logins, cookies and VPN, while you keep your screen, mouse and keyboard. It is a local stdio server with 11 tools, from a one call task loop down to single verified clicks.

Every action is verified against the window state afterwards. A click that lands nowhere is reported as a failure, never as a silent success.

Full documentation: https://agentfenster.com/mcp

The MCP server ships with the agentfenster app, a commercial closed-source Windows product. This repository is its public documentation and issue tracker. It contains no product source code.

Install

  1. Install agentfenster on Windows 10 or 11 from https://agentfenster.com. The MCP server is part of the app. Every plan starts with a 3-day trial, no card and no account required.
  2. Register the server with Claude Code:
bash
claude mcp add agentfenster -- agentfenster-mcp

Any MCP client that can launch a local stdio server works the same way. Then just ask:

> open the character map on the agent desktop and turn on the advanced view

The 11 tools

ToolWhat it does
list_windowsLists the windows this machine can drive, with title, class, minimized state and whether a picture can be captured. The starting point when you do not know the exact window title.
desktop_taskRuns a full GUI task in one window and returns a single paragraph: what happened, whether it worked, and if not, which step failed and why. The click loop runs inside the server, so its retries never enter your context.
inspect_windowReturns the capped, filtered element tree of one window as a compact list, for planning clicks yourself. Capped at 200 elements and never walked from the desktop root.
replayLists recent runs, or returns one run's step-by-step log: every plan, every action, whether it changed anything, and a link to the visual replay in the app.
open_appLaunches a program on the shared agent desktop and reports its first titled window. The agent desktop starts empty, so this is how anything gets onto it.
read_desktopReads the compact element tree of a window on the agent desktop, one line per interactive element, no screenshot. Plan your clicks from this.
clickClicks one element, chosen by id or name from read_desktop, then verifies against the desktop state and states explicitly whether anything changed.
type_textTypes text into one element on the agent desktop, then verifies whether anything changed.
press_keyPresses a key or key combination such as ctrl+s, enter, tab or alt+f4, optionally focusing an element first, with verification afterwards.
scrollScrolls a window or a specific pane and reports whether the content actually moved. A list that is already at the bottom reports no effect.
agent_windowsLists the windows currently open on the shared agent desktop, frontmost first, so you can pick one by a title substring.

How it works

  • The agent works on a second Win32 desktop inside your session. It never takes your foreground, mouse or keyboard, and you can watch it live in the app.
  • Tools read the UI Automation element tree instead of screenshots: one line per interactive element, capped at 200 elements per window.
  • The click loop of desktop_task runs inside the server on a model endpoint you configure, so its retries and intermediate steps never enter your context.
  • Risky actions such as delete, send, pay or install are denied by default on MCP calls and reported, instead of blocking on a confirmation nobody can answer.
  • agentfenster sends no telemetry. The only network traffic is the model endpoint you configure for the click loop.

Requirements

PlatformWindows 10 or 11
Transportstdio, local
RuntimePython 3.11 or newer, installed with the app
Prerequisitethe agentfenster app (commercial, 3-day trial)

Links

Issues

Use the issue tracker of this repository for bug reports and questions about the MCP server: tool behaviour, install problems, MCP client compatibility. For anything about the app itself, Discord and [email protected] are faster.

Rights

agentfenster is commercial closed-source software. No open-source license is granted here. The text in this repository is published so that MCP users and directories can read, quote and link to it. Use of the agentfenster app itself is governed by the terms at https://agentfenster.com/terms.

Source: README.md at commit f24415b

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.0LatestSep 16, 2026

Score

40% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts

    Declare a license in the repository or the registry entry.