Security
Last updated: October 11, 2026
The short version
- The agent works on a second desktop inside your own Windows session. Nothing runs in our cloud.
- Credentials live in the Windows Credential Manager. The model only ever sees placeholders.
- Every outgoing connection is listed below. There is no other telemetry.
- Updates are only installed if their Ed25519 signature checks out against a key built into the app.
Data flow
The agent reads a window’s UI element tree (and, for canvases and games, a picture of that one window), decides the next step with the model you configured, and acts through the window itself. It never sends real mouse or keyboard input to your screen. Runs are recorded locally as replays in %LOCALAPPDATA%\agentfenster; they never leave the PC.
Stored credentials and 2FA codes are filled in locally at the moment of use, only into a window that matches the app you stored them for. Every log line and every report runs through the same masking first, so a vault value that shows up in a window title becomes a placeholder.
Every network connection
Every agent step
To: The model endpoint you configured (Claude Code, Codex, an API, or a local model)
Sends: The task, the element tree of the target window and, where needed, a window picture. Vault values are replaced by placeholders before they leave.
After a run, if the report switch is on and a key is activated
To: Our server, 91.98.138.20:8443
Sends: Metadata only: version, Windows version, hashed device id, licence tier, step count, duration, tokens, success, error kinds.
Key activation, then about once a month
To: Our server, 91.98.138.20:8443
Sends: The licence key and the hashed device id.
Update check
To: Our server, 91.98.138.20:8443
Sends: A request for the signed manifest of your channel. No id, no key.
Our server presents a self-signed certificate that the app pins by its public-key hash, so a network in between cannot read or alter the traffic. Pin (SPKI, SHA-256, Base64):
VU3sAPrQoCUad1aagCtWbczbHo5klbgLHl0MZCCKGoo=The local interface is a server on 127.0.0.1 that refuses any request not coming from this machine, checks Host and Origin, and requires a per-process session key. Verify all of this yourself with any packet capture, for example Wireshark or pktmon.
Signed updates
The app downloads an update only after the manifest’s Ed25519 signature verifies against this public key, built into every install. The installer’s SHA-256 from the signed manifest must then match the downloaded file. Older manifests are refused, so nobody can push you back to an earlier version.
ef84b2be6b06698ac7553f6cb1f2c4551f1b97c8181630d50f260f64fe6cf699Verify a download
Every public release lists the SHA-256 of its installer here and in the changelog. Check yours in PowerShell before you run it:
Get-FileHash .\agent-hub-setup.exe -Algorithm SHA256There is no public release yet, so there is no hash to compare against. This list fills with the first download.
The installer is not code-signed yet. Windows SmartScreen asks once before it runs, and on a PC with Smart App Control turned on Windows refuses unsigned programs entirely. Defender may also flag the hidden-desktop feature, because the same Windows API is used by some malware; the FAQ on the home page explains the difference.
What we do not defend against
- Text on a screen is untrusted input. A web page or document can contain text written to look like an instruction. The permission switches (buy, send, delete, install, real mouse) are off by default for this reason; keep them off for unattended runs.
- A program already running as you can do what you can do. The local interface stops other machines and web pages, not malware on your own account.
- The clipboard is shared with Windows: a copy by the agent overwrites yours.
Report a vulnerability
Write to [email protected] with “Security” in the subject. We answer within three working days, fix confirmed issues before anything else, and credit you in the changelog if you want.