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.
Agent
The agent works on its own desktop. Watch it, step in, or take over.
Simulated session for illustration
A computer-use agent for Windows. It drives real Win32, WinForms, WPF, Office, Chrome, Edge and Electron apps.
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.
Simulated session for illustration
Supervise up to 6 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.
Simulated session for illustration
Your terminal, in the same window.
A real ConPTY terminal with an MCP server built in — run Claude Code, Codex CLI or Gemini CLI in it with your own sign-in. The agent and your terminal share one desktop.
Terminal
A shell in this window: run Claude Code next to the agent.
Simulated session for illustration
Run started
run 014: invoice export
Click “File”
Click “Export as PDF”
Click “Save”
no effect detectedREPORTED AS ERROR
Retry: click “Save”
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
CreateDesktopAPI. - The agent reads the
UIAelement 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
ConPTYterminal plus MCP server, for Claude Code, Codex CLI or Gemini CLI. - 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 account. Session reports are opt-out, not silent.
- Memory remembers what worked in an app: one note per lesson, searched full text, never one copy per run.
- A schedule can trigger a run on time, on a new file, or on app start, and refuses any step that deletes, sends or pays.
- Preview: a phone on the same Wi-Fi can send a task and watch it run.
- Preview: say the task instead of typing it. Transcription runs on this machine, nothing leaves it.
- The app checks a signed update manifest and installs the new version itself.
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, plus a small, masked session report you can switch off in Settings — never screenshots, element trees or vault secrets (see /privacy).
- Verify it yourself with any packet capture.
What each feature looks like
Real screenshots from the app, one state per feature. No renders.
Voice
Speak the task instead of typing it. What was heard shows above the task bar; nothing runs until you press Enter yourself.
Phone
Turn on “Let a phone watch this machine”, then scan the QR code: a phone on the same Wi-Fi can watch a session and answer for it.
Memory
Search what the agent learned from earlier runs, grouped by app, or look at it as a graph of what connects to what.
Skills
Show the agent a task once and it stores your intent, not your pixels; this is the recording view while it is capturing your steps.
Routines
A schedule can trigger a run on time, on a new file, or on app start; here it repeats once an hour.
What it cannot do
These are measured limits of Windows, not missing features. The same list ships inside the app.
- Windows has no API to move a window between desktops. The agent hands over the result instead: a file, a URL, text.
- There is no real mouse pointer on the agent desktop, and raw input is blocked there by Windows itself. On your own screen it still never sends a real click or keystroke, by choice, not by limitation.
- Store apps Windows itself activates, Calculator, the Microsoft Store, Camera, Clock, Sound Recorder, WhatsApp Desktop, Microsoft Teams, always land on your own desktop. The name is recognised and the start refused before a window ever opens.
- Chromium apps (VS Code, Discord, Slack, Obsidian) need a debugging port, and only get one if agentfenster starts them itself.
- The agent browses in its own Chromium profile, so it starts signed in to nothing. Sign in once yourself, on your own screen, and it keeps the session from then on.
- The clipboard is shared with Windows: a copy by the agent overwrites yours.
- No CAPTCHA solving. The agent may click a plain confirmation checkbox; an image puzzle pauses the run for you to solve.
- Games, DirectX, Citrix or RDP, Adobe canvases, Java Swing, elevated processes and DRM content are out of reach for any accessibility based tool.
- The installer is unsigned for now: Windows SmartScreen, and sometimes Defender, flag it on first run, the same way they flag any new, unsigned program.
What it will cost
Planned launch pricing. agentfenster is in development and not for sale yet.
Free
your own visible desktop, forever
Everything that reads or works on the desktop you are looking at. The app itself, the live view and every replay stay free for good.
- Read any window as an element tree
- Hand a goal to a window you opened yourself
- Replay recorded runs
- Live view, take-over, recordings, terminal
Planned for launch. Prices may still change.
Agent desktop
the invisible second desktop, unlocked
or $89 a year, two months free
The agent works on a desktop you never see while you keep yours: it opens apps, clicks, types, scrolls and takes whole goals, on up to 6 agent desktops at once.
- Open and drive apps on the invisible agent desktop
- Up to 6 agent desktops in parallel
- Teach by doing, replay without a model
- Credential vault and 2FA mailbox
- 14-day trial of everything: no key, no card, no account
Planned for launch. Prices may still change.
Founder price: the first 100 customers keep $9 a month or $89 a year for good, even when the list price rises later.
Why the price is low
Our costs are low, so the price is.
The agent reads the UI element tree instead of streaming screenshots to a vision model.
A task you recorded once replays without any model at all.
Claude Code runs through your own access, so the subscription pays for the software.
Frequently asked
Can I use it today?
Isn't a hidden desktop what malware does?
Who sees my credentials and 2FA codes?
Does anything leave my machine?
Does it take over my mouse or keyboard?
What happens when a click has no effect?
Is this available in Europe?
Why is the price this low?
Why might Defender flag the installer?
Give your agentits own desktop
agentfenster will be a paid product: your subscription is the business model. There is no data to monetize, no account, and no report you can't switch off.