GUI agents on Windows leave you with two bad options today: either the agent takes over your screen and mouse, or it sits in a cloud VM where none of your logins, cookies, VPN or two-factor codes exist. agentfenster does neither.
It creates a second Win32 desktop inside your running Windows session: same user, same credentials, same browser cookies. That desktop is never displayed, so the agent cannot grab your foreground; you watch it as a live stream in the window and step in whenever you want. Six such desktops run side by side by default, adjustable up to 16.
The second point is honesty: every action is verified against the window state afterwards. If nothing changed, the step counts as a failure, not a silent success you notice three days later.
Recurring tasks you demonstrate once yourself. agentfenster writes along: clicks, typing, accessibility context, and replays the run afterwards with no model at all, which makes it free and reproducible. Every run can be replayed as a film with a timeline and exported as MP4. Credentials live in the Windows Credential Manager and the model only ever sees a placeholder; buying, registering, sending, deleting and installing are individual switches, all off by default.
Claude Code runs in a real terminal in the same window and drives the same desktop through the bundled MCP server with 13 tools. The project is pre-launch; the MCP documentation and the product page are public, the application code stays closed.