Spin up a real Linux desktop in a disposable microVM, publish the VNC port, and get a live preview URL your agent can watch and control. Screenshot, click, type — then throw it away.
Full Linux desktop streamed over VNC · disposable microVM · $5 free, no card
$ pip install pilldrops Successfully installed pilldrops-1.4.0 $ python -c " from pilldrops import Sandbox with Sandbox.create(template='desktop', ports=[5900]) as sb: print(sb.url)" https://5900-sb-9f2c1a.pilldrops.app ✓ Linux desktop live over VNC · disposable microVM
The desktop template boots a full Linux graphical environment inside an isolated sandbox. Publish the VNC port and you get a preview URL you can open in a browser or feed to a vision model.
A real graphical session — window manager, cursor, keyboard — running in the sandbox, not a fake DOM. Launch it with template="desktop".
Publish port 5900 and sb.url returns a preview URL. Open it to watch your agent move the mouse and type in real time.
Run an actual browser, file manager, and text editors inside the desktop — the same apps a human operator would drive.
Each desktop is a throwaway microVM. Kill it when the task ends and every trace goes with it — no cleanup, no shared state.
The pattern is simple: capture the screen, let your model decide, then send a click or keystroke back into the desktop. Below is the shape — the full loop, including how to wire screenshots into a vision model, lives in the guide.
from pilldrops import Sandbox with Sandbox.create(template="desktop", ports=[5900]) as sb: print(sb.url) # watch the desktop live over VNC # 1. perceive — grab the current screen sb.run("scrot /tmp/screen.png") img = sb.read("/tmp/screen.png") # feed to your vision model # 2. decide — your model returns an action (click/type) # action = model.decide(img) -> {"x": 420, "y": 300} # 3. act — send the input back into the desktop sb.run("xdotool mousemove 420 300 click 1") sb.run("xdotool type 'hello from the agent'")
Wiring screenshots into a model, handling coordinates, and running the loop to completion is covered end to end in the computer-use guide.
The guide walks the entire flow — launching the desktop template, opening the VNC preview, feeding screenshots to a Claude computer-use style model, and sending clicks and keystrokes back — with code you can run on $5 of free credits.
Step-by-step: create the desktop sandbox, get the preview URL, capture the screen, and drive the mouse and keyboard from your agent — the honest, working loop end to end.
with Sandbox.create(template="desktop", ports=[5900]) as sb: print(sb.url)
Prefer the CLI? Create the same desktop sandbox with the VNC port published in one command, then grab the preview URL from the output.
$ pilldrops sandbox create \
--template desktop --port 5900
A real Linux desktop over VNC, billed per second, disposable when the task ends. Launch one on $5 of free credits — no card — and follow the computer-use guide for the full loop.