Most AI agents run on your own machine. I wanted mine to have a computer of its own: one it can break, bill by the second, and throw away.
So I forked nanoMuse, an open-source personal agent, and built a cloud PC toolset for it. In the v1.0.9 release the agent can:
- provision a cloud machine on Daytona or InstaBox
- run commands, move files, take screenshots
- operate a real graphical desktop (click, type, open pages)
- destroy the machine when the job is done
Two design choices I care about:
A provider protocol, not a vendor hack. One CloudProvider interface, two implementations. The desktop on InstaBox does not use VNC at all: it rides the exec channel with DISPLAY=:1, xdotool and scrot, which I verified by reading the official TypeScript SDK.
Cost safety is a feature, not a hope. Every action passes a safety approval layer. Spend is written to a local ledger at published rates and reconciled against vendor metering (0.31% error against InstaBox per-second usage in my test). A monthly budget gate can block new machines before any provider call, and an idle sweeper destroys machines that sit unused, so nothing bills by accident.
The cloud PC runs on your own free-tier provider account (both vendors work without a card). The software ships no machines.
Downloads for Windows, Android, macOS and Linux:
https://github.com/Thanksgiver233/nanoMuse/releases/tag/v1.0.9
Code (GPL-3.0): https://github.com/Thanksgiver233/nanoMuse
If this is useful to you, a star on GitHub helps a lot. Issues and feedback welcome.
Top comments (0)