Desktop & computer use
Every VM can run a full XFCE desktop with Chromium and Firefox. Humans watch it live in the browser. Agents drive it through a screenshot and input API. Pause it, and the windows are still there when it wakes.
People need a live stream. Agents need pixels and an input channel, which is far cheaper. We do both.
Open https://<short_id>-desktop.taskhost.cloud in Chrome, Firefox or Safari. Our target is a visible desktop in under 2 s from the click, including waking a paused VM. KasmVNC runs over WebSocket on HTTPS: no UDP, no TURN server, no client install.
GET /screenshot returns a PNG (target p95 under 300 ms). POST /input moves, clicks, scrolls, types and sends key combos. The shape matches Anthropic's computer-use tool, so existing agent loops plug in directly.
Computer use API
Two endpoints are enough for most agent loops: take a screenshot, decide, send input, repeat. No video stream to decode.
# Take a screenshot
curl -H "Authorization: Bearer $TASKCLOUD_API_KEY" \
https://api.task.cloud/v1/sandboxes/$ID/screenshot -o screen.png
# Click, then type
curl -X POST -H "Authorization: Bearer $TASKCLOUD_API_KEY" \
-H "Content-Type: application/json" \
https://api.task.cloud/v1/sandboxes/$ID/input \
-d '{"actions":[
{"action":"left_click","coordinate":[640,410]},
{"action":"type","text":"hello from the agent"},
{"action":"key","text":"Return"}]}'import { Client } from "@taskcloud/sdk";
const client = new Client(); // reads TASKCLOUD_API_KEY
const sbx = await client.sandboxes.create({ template: "desktop" });
const shot = await sbx.desktop.screenshot(); // { data (base64), media_type, width, height }
await sbx.desktop.input({
actions: [
{ action: "left_click", coordinate: [640, 410] },
{ action: "type", text: "hello from the agent" },
{ action: "key", text: "Return" },
],
});
// Let your user watch, read-only, for 10 minutes
const { url } = await sbx.desktop.stream({ ttl_seconds: 600, read_only: true });The pause snapshot includes RAM, so windows, the open browser and logged-in tabs return in under a second after resume. The stream reconnects automatically.
Ubuntu 24.04, XFCE, KasmVNC, Chromium, Firefox, xdotool and fonts. The template itself is a RAM snapshot of a running desktop, so new VMs start on a ready screen.
Viewer links are signed and expire (10 minutes by default). Share read-only views with your users and revoke links through the API. The VNC port is never exposed without a token.
An active viewer counts as activity, so the VM stays up while someone watches. An idle desktop with no viewers and no input pauses like any other VM.
Screen encoding uses the VM's own vCPUs, with no extra fee. That is why desktops need at least default (4 vCPU / 8 GB).
Rendering is software-only. Office work, browsing and dev tools are fine; WebGL, 3D maps and video playback are slow. Smooth 60 fps WebRTC streaming is on the roadmap, not in the beta.
Desktop streaming and computer use are part of the beta, on every default or large VM, at the same per-second price.