Skip to main content

Getting started

Last updated: October 11, 2026

1. What you need

  • Windows 10 or 11, 64-bit. No admin rights needed.
  • Claude Code or Codex CLI, installed and signed in. Your existing subscription does the thinking; agentfenster does not resell model access.
  • About 500 MB of free disk space.

2. Install

Download agent-hub from the download page, run it and pick agentfenster. It installs per user into %LOCALAPPDATA%\Programs\agentfenster, adds agentfenster-mcp to your PATH and starts the 14-day trial of everything. The installer is not code-signed yet: SmartScreen asks once, choose More info and then Run anyway. Compare the file hash first if you like, as shown on the Security page.

3. Open a new terminal

Close your terminal and open a new one. A terminal only reads PATH when it starts, so the one you had open before the install cannot find agentfenster-mcp yet. This is the most common reason the next step fails.

4. Connect your coding agent

With Claude Code, in the new terminal:

claude mcp add agentfenster -- agentfenster-mcp

With Codex CLI:

codex mcp add agentfenster -- agentfenster-mcp

To share the setup with a project instead, put this .mcp.json in the project folder (Claude Code picks it up and asks once whether to trust it):

{
  "mcpServers": {
    "agentfenster": {
      "command": "agentfenster-mcp",
      "args": []
    }
  }
}

For Codex the same goes into ~/.codex/config.toml:

[mcp_servers.agentfenster]
command = "agentfenster-mcp"

Still “command not found”? Use the full path instead:

claude mcp add agentfenster -- "%LOCALAPPDATA%\Programs\agentfenster\agentfenster-mcp.exe"

Check it worked with claude mcp list: agentfenster should show as connected.

5. Your first task

Start Claude Code or Codex and ask, in plain words:

open the character map on the agent desktop and find the euro sign

The agentfenster app opens and shows the agent desktop live. You keep working on your own screen; your mouse is never touched. Every step is checked for effect, and the run is kept as a replay under History.

The first start of the app asks three things: a quick look at the agent desktop, which AI it should use, and one harmless test task. It also shows the limits Windows sets, worth one minute of reading.

6. Where to go next