1. The AI, tool by tool — and for real
NodalDesk lets you authorize an AI to see, hear and operate your machine through its own channel, with your permission and everything logged. In the panel, Agents section, you expose 13 tools (read screen, screenshot, video, monitors, mouse, keyboard, type text, scroll, hear…). The 0.2.5 news: all 13 switch off for real.
Before, two of them (the live video and the monitor list) followed the general
screen permission instead of their own switch. Not anymore: turn frame
off for an agent and the host closes its video channel; turn monitors
off and it gets no monitor list and can't switch monitors; turn
read_screen off and it can't read the accessibility tree, not even
through the tunnel. Each switch cuts the real capability, not just the label.
How to test it: connect your AI, go to Agents, turn a specific tool off and ask it to do that action. It won't — and the host's activity log shows the rejection. The point: you're in charge, per tool, locally, and you can revoke whenever you want.
2. One state per node — and the path to fix it
Before, the same machine could show up "active" in one view and "inactive" in another, because each screen mixed two sources (what the cloud sees and what happens on the machine itself). Now there's one truth of state per node, shared by web, desktop and Android.
And when a node needs attention, it says so with its cause and its path: if its credential was revoked it shows "needs re-linking" (open it on that machine), if it's sleeping it offers to wake it, if your session lost the cloud it tells you to reconnect. Old alerts now expire too: no more "ghost" alerts coming back for no reason.
3. Windows is a host, not just a client
We say it plainly because the old material didn't: on Windows, NodalDesk works
as a host (the machine you control) as well as a client. There's
an .exe installer and a portable build, with native screen capture and
keyboard/mouse injection. macOS, for now, is a client. Linux, Windows and Android
are real hosts.
4. Steadier, with direct billing
The Nodes section stopped overloading: we closed a request storm that could fire 429 errors under normal use. And from desktop and Android, managing your plan now opens Stripe directly with your session, no bouncing through the web panel.
The limits, no make-up
- Connecting a new AI still goes through the SDK/MCP: the channel, permissions and monitor are live, but there's no "connect your AI" button for everyone yet. Continuous autonomous control is the next step, not a finished feature.
- Wake-on-LAN is honest beta: it wakes a machine if you keep another node powered on its same local network and the machine has WoL enabled in the BIOS. It's not magic "wake over the internet", and it still needs deeper validation against a truly powered-off machine.
- Control from boot (choosing the OS on restart) has its Linux engine done, but it's missing UI and physical dual-boot testing: we don't call it finished yet.
How to update
The desktop updates itself (auto-updater). If you'd rather download by hand, the
0.2.5 set is on Downloads: an AppImage and a
.deb installer for Linux, installer and portable for Windows, and the
Android APK. The account and backend are already up to date.
What's next
Keep refining the AI experience (a simpler path to connect it), finish the control-from-boot UI with real validation, and polish video definition and smoothness over relay. We'll tell you when it's done and tested, not before.