Skip to main content
In development

Your agent getsits own desktop

A hidden second desktop inside your Windows session: same logins, same VPN, same MFA. You watch live and step in anytime. It never takes your mouse.

Watch the live demoJoin the Discord
No release yet. Join the Discord to follow the build and be first in.
agentfenster

Agent

The agent works on its own desktop. Watch it, step in, or take over.

Live
Calculator
12
%
CE
C
⌫
7
8
9
×
4
5
6
−
1
2
3
+
±
0
,
=

Simulated session for illustration

Drives real Windows apps, not only the browser.

A computer-use agent for Windows. It drives real Win32, WinForms, WPF, Office, Chrome, Edge and Electron apps.

Win32WinFormsWPFOfficeChromeEdgeElectron

The whole tool, tab by tab

It works. You watch.

The agent runs real apps on its own desktop. Here it fills a customer form field by field. You see every move as a live stream with a virtual cursor, and take over the moment you want to.

Agent

The agent works on its own desktop. Watch it, step in, or take over.

Live
Customer Records

New customer

Full nameAlex Meyer
Email[email protected]
CompanyAcme Logistics
CancelSave

Simulated session for illustration

Supervise up to 8 desktops from one screen.

Each agent gets its own hidden desktop in your session. One view, all runs, step in anywhere. Your own mouse and windows stay untouched.

Agent desktops

Run several agents side by side, each one on its own invisible Windows desktop, none of them on yours.

Open desktops

Up to six at a time, each one a real Windows desktop of its own.

Simulated session for illustration

Teach it once. It replays without a model.

Show the agent a task by doing it yourself. It stores your intent, not your pixels, and replays it with zero model calls.

As of August 2026, OpenAI ships record-and-replay on macOS only and excludes the EEA, the UK and Switzerland. agentfenster records and replays on Windows and runs locally in Europe.

Recordings

Every run is kept as a replayable film. Scrub it, step through it, change the speed.

Replay
Calculator
12
%
CE
C
⌫
7
8
9
×
4
5
6
−
1
2
3
+
±
0
,
=
18.0s
18.0s

Simulated session for illustration

Claude Code, in the same window.

A real ConPTY terminal with an MCP server built in. Your Claude setup, your keys: the agent and your terminal share one desktop.

Terminal

A shell in this window: run Claude Code next to the agent.

Windows PowerShell

PS C:\Users\you>

Simulated session for illustration

0:02

Run started

run 014: invoice export

0:07

Click “File”

0:11

Click “Export as PDF”

0:16

Click “Save”

no effect detectedREPORTED AS ERROR

0:19

Retry: click “Save”

0:24

Done

invoice.pdf saved, every step verified

Every action is checked for effect

Each click is checked against the window's element tree. No change means an error: reported, never dressed up as success.

What it actually is

  • A second Win32 desktop inside your Windows session, created with the standard CreateDesktop API.
  • The agent reads the UIA element tree, not screenshots.
  • Live stream with a virtual cursor; take over anytime. It never takes your mouse.
  • Every run is stored as a replay with a timeline.
  • Real ConPTY terminal plus MCP server for Claude Code.
  • Credentials stay in the Windows Credential Manager: the model sees placeholders.
  • 2FA codes are read from your Gmail or Outlook inbox, locally.
  • Runs locally. No telemetry. No account.

What it can see, and what it can't

  • The model sees element trees and placeholders, never your stored credentials.
  • Network traffic consists of the model calls you configure; there is no telemetry endpoint.
  • Verify it yourself with any packet capture.

What it will cost

Planned launch pricing. agentfenster is in development and not for sale yet.

Cockpit

one agent desktop

$4/month

One agent desktop with every core feature: live view and take-over, replays, Claude Code over MCP, teach by doing.

  • 1 agent desktop
  • Live view and take-over
  • Replays of every run
  • Claude Code over MCP
Get notified on Discord

Planned for launch. Prices may still change.

Fleet

up to 8 agent desktops in parallel

$7.49/month

Everything in Cockpit, scaled up: run agent desktops in parallel, plus the 2FA mailbox and priority support.

  • Up to 8 agent desktops in parallel
  • Everything in Cockpit
  • 2FA mailbox
  • Priority support
Get notified on Discord

Planned for launch. Prices may still change.

Why the price is low

Our costs are low, so the price is.

element tree, not screenshots

The agent reads the UI element tree instead of streaming screenshots to a vision model.

replay: 0 model calls

A task you recorded once replays without any model at all.

your own Claude setup

Claude Code runs through your own access, so the subscription pays for the software.

Runs entirely on your machine, no account required

Frequently asked

Can I use it today?
Not yet. agentfenster is in active development: there is no download and nothing to buy. The demo on this page shows the real interaction model. Development happens in the open on our Discord: join to watch live runs and be first to know when the trial opens.
Isn't a hidden desktop what malware does?
Hidden desktops are a known malware technique: MITRE tracks it as T1564.003. The underlying Win32 API, CreateDesktop, has been a standard Windows feature since Windows 2000. The difference is direction. Malware hides a desktop from you and streams it to an attacker. agentfenster hides the desktop only from your foreground and streams every frame to you. You watch live, you can take over at any point, and every run stays on your machine as a replay. No remote access, no persistence outside your own session.
Who sees my credentials and 2FA codes?
Credentials are stored in the Windows Credential Manager. The model only ever sees placeholders: real values are filled in locally at the moment of use. 2FA codes are read from your Gmail or Outlook inbox on your machine and handled the same way.
Does anything leave my machine?
The model calls you configure are the network traffic the app initiates. There is no telemetry, no account, no analytics endpoint. Verify it yourself with any packet capture.
Does it take over my mouse or keyboard?
No. The agent works on its own Win32 desktop inside your session. Your mouse, keyboard and windows stay yours. You watch the agent's desktop as a live stream and take over whenever you choose.
What happens when a click has no effect?
Every action is checked against the window's element tree afterwards. If nothing changed, the step is reported as an error: in the live view and in the replay timeline. It is never recorded as a success.
Is this available in Europe?
Yes. Everything runs locally on your Windows machine; there is no regional gating. For context: as of August 2026, OpenAI's record-and-replay ships on macOS only and excludes the EEA, the UK and Switzerland, per OpenAI's own documentation.
Why is the price this low?
The agent reads the UI element tree instead of streaming screenshots to a vision model, so each step costs a fraction of what screenshot-based agents spend. Recorded tasks replay without any model at all. Claude Code runs through your own setup: the subscription pays for the software, not for compute.
Why might Defender flag the installer?
Windows Defender's reputation heuristics sometimes flag new, low-distribution executables. Every release lists its installer's SHA256 hash so you can verify exactly what you downloaded. If Defender flags a verified download, it is a false positive: report it to Microsoft or contact us and we will resubmit the build.

Give your agentits own desktop

Join the DiscordNo release yet. Join the Discord to follow the build and be first in.

agentfenster will be a paid product: your subscription is the business model. There is no data to monetize, no telemetry, no account.