Skip to main content
MCP server for Claude Code and Codex

Give your coding agent its ownWindows desktop

Your agent drives real Windows apps on a hidden second desktop inside your session: same logins, same VPN, same MFA. You watch live and step in anytime. It never takes your mouse.

Join the waitlistWatch the demo
Release candidate in testing. No payment taken: we email you once when it opens.
agentfenster

Live

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

Live
Character Map
Font:Arial
!
"
#
$
%
&
'
(
)
*
0
1
2
3
4
5
6
7
8
9
A
B
C
D
E
F
G
H
I
J
¡
¢
£
¤
¥
¦
§
¨
©
ª
«
¬
®
¯
°
±
²
³
´
µ
Characters to copy:©®Select

Simulated session for illustration

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.

Live

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 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

Every run is kept as a film.

Scrub a timeline of every step and whether it changed anything, or export the run as an MP4. A run that reached its goal is saved as a skill: ask again and it replays with zero model calls.

Recordings

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

Replay
Character Map
Font:Arial
!
"
#
$
%
&
'
(
)
*
0
1
2
3
4
5
6
7
8
9
A
B
C
D
E
F
G
H
I
J
¡
¢
£
¤
¥
¦
§
¨
©
ª
«
¬
®
¯
°
±
²
³
´
µ
Characters to copy:©®Select
23.4s
23.4s

Simulated session for illustration

Driven from the terminal you already use.

Claude Code or Codex runs in your own terminal, with your own sign-in, and drives the agent desktop over MCP. Every tool call checks its effect before it answers.

MCP

Claude Code or Codex in your own terminal drives the agent desktop.

Your terminal, your sign-in

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.

A real run, start to finish

One uncut recording of the agent doing a real task on its own desktop, while the screen in front of it stays free.

Video in production

We are recording a real, uncut run for this spot. Until it is here, the demo at the top of the page shows the interaction model as a simulation.

Measured against Windows‑MCP

Windows-MCP 0.8.5, the most used open Windows automation server. Seven A/B/A rounds on the same window and the same visible desktop, 30 August 2026. The verdict comes from the window itself, a file it writes and its text, never from either tool's own report.

One look at a windowagentfenster254 tokens, 36 msWindows-MCP7,066 tokens, 1,008 ms
One clickagentfenster24 msWindows-MCP565 ms
Clicks on a button covered by another window that workedagentfenster14 of 14Windows-MCP0 of 7
Clicks reported as success with no effectagentfenster0 of 14Windows-MCP7 of 7
Actions that moved your mouse pointeragentfenster0 of 28Windows-MCP14 of 14

Where we do not win

  • With nothing in the way, both reach 100 % (28 of 28 and 14 of 14).
  • Typing into a field is slower on our side: 2.5 s against 1.5 s, not yet diagnosed.
  • Tree against tree is only 2.1× smaller. Most of the 27.8× on a look comes from Windows-MCP reading the whole desktop (264 elements) where we read one window (8).

And screenshot computer use?

Computer use in Claude and Codex looks at screenshots, and on Windows it works on the screen in front of you, with your mouse. agentfenster reads the element tree on a desktop you never see. Same task on the Character Map, covered by another window: 3 steps in 1.5 s, 670 tokens on the tree against 4,680 on our own screenshot path.

Numbers from one machine and one test window; your apps will differ. Method and raw data on request.

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.
  • An MCP server for Claude Code and Codex: your own CLI, your own sign-in, nothing resold.
  • Interface in English and German. It follows your Windows language.
  • 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.
  • The agent keeps notes on what worked in each app: one note per lesson, searched full text, never one copy per run.
  • A run that reached its goal becomes a skill and replays with zero model calls. When a step no longer fits, the model takes over from exactly there.
  • The agent has its own browser profile, driven over the Chrome DevTools Protocol. Every click and every typed field is read back from the page.
  • A frozen app is reported as not responding. It never freezes the agent with it.
  • The app checks a signed update manifest and installs the new version itself, on the Stable or the Alpha channel.

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, the licence and update checks, and a small session report of counters you can switch off in Settings.
  • That report never carries screen content, window titles or vault secrets (see /security).
  • Verify it yourself with any packet capture.

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 that Windows itself activates always land on your own desktop: Calculator, the Microsoft Store, Camera, Clock, Sound Recorder, WhatsApp Desktop, Microsoft Teams. 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 profile, so it starts signed in to nothing. A login is handed to you: with your OK the same profile opens on your screen, you sign in and close it, and the agent carries on invisibly with that session.
  • 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.

Pricing

Launch pricing. agentfenster is not on sale yet: join the waitlist and we email you once when it is.

Free

your own visible desktop, forever

$0no key, no clock

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, history
  • MCP for Claude Code and Codex

Included in every install. No key needed.

Pro

the invisible agent desktop

$9/month

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
  • Finished runs replay without a model
  • Credential vault and 2FA mailbox
  • One key for up to 3 of your PCs
  • 14-day trial of everything: no key, no card, no account
Join the waitlist

No payment is taken. Prices may still change before launch.

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.

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 that worked 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 on your machine. No sign-up, no login, no profile: buying sends you a licence key by email.

Licence, trial and refunds

How does the 14-day trial work?
Every install starts with 14 days of everything, including the agent desktop. No key, no card, no email. When the trial ends nothing is deleted and nothing stops working: the app falls back to Free, so you can still watch, replay and work on your own desktop.
What exactly do I pay for?
The invisible agent desktop: the agent opening and driving apps where you cannot see them, up to 6 at once, plus the credential vault, the 2FA mailbox and replays without a model. Reading windows and working in windows you opened yourself stay free.
How does the licence key work?
After checkout you get one key by email. Enter it in Settings and it covers up to 3 of your own PCs; you can free a PC again yourself. The key is checked once when you activate it and about once a month after that. Without a key, the app never calls our licence server.
Can I cancel, and can I get a refund?
You can cancel anytime; the plan then runs to the end of the period you paid for. Every buyer, consumer or business, can ask for a full refund within 14 days of the first payment, no reason needed. Write to us with your order and we refund it to the original payment method. Details are on the refund page.
Who handles the payment?
Polar (polar.sh) acts as merchant of record: it sells the licence, takes the payment, charges the right VAT for your country and sends the receipt. The software itself is licensed by Ilse Grabner, Austria. We never see your card details; we only receive the email address that the licence key goes to.

Frequently asked

Can I use it today?
Not publicly yet. agentfenster 0.4.0 is a release candidate, tested by a small group. There is no public download and nothing is for sale. Join the waitlist and we email you once when it opens; to watch the build, join the Discord.
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 app talks to the model endpoint you configure. It also talks to the agentfenster server in three cases: when you activate a licence key (and about once a month after), when it checks for updates, and for a small session report after each run (app version, step count, duration, success; never screen content, window titles or vault secrets). The report is on by default so we can find bugs, and it is one switch away from off in Settings. The full list is on /privacy and /security. 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.
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. A task that worked once replays 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 waitlistRelease candidate in testing. No payment taken: we email you once when it opens.

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.

Join the waitlistWatch the demo