Guide 04 · template: desktop
Computer-use agents on a real Linux desktop
Computer-use models (Claude computer use, OpenAI CUA) click and type on real
software. The desktop template boots a full Linux desktop streamed over VNC —
screenshot it, feed the frame to the model, execute its clicks.
1
Boot the desktop, expose the VNC port
$ pilldrops sandbox create --template desktop --port 5900 \ --vcpu 2 --mem 4096 --timeout 7200 sb_d3sk10 running desktop app_url: https://5900-sb_d3sk10.…
2
Watch it live
Open the preview URL in any noVNC client (or an iframe in your product) — you see exactly what the agent sees, in real time.
3
Drive it from your agent loop
# screenshot → model → action, the standard computer-use loop sb.run("import -window root /tmp/screen.png") # capture frame = sb.files.read("/tmp/screen.png") # → model sb.run(f"xdotool mousemove {x} {y} click 1") # act
≈ $0.17/h (2 vCPU · 4 GB) ·
preview URLs