Computer-use API
The actions computer-use models emit, as they emit them.
- Screenshot, click, drag, scroll
- Key, type, clipboard, open URL
- wait_for_screen_change
A Linux desktop with Chrome, per agent. Your agent drives it with screenshots, clicks and keys. Your users watch it in your app and take over when a person is needed.
$0.04 per vCPU-hour·Embeddable live view·Human handoff
The actions computer-use models emit, as they emit them.
The agent stops and asks for a person. Your user presses Take control, signs in, presses I'm done, and the agent goes on. Agent input is refused meanwhile.
A signed link for an iframe. Nothing of ours on it. Watch only, or with control. Events and commands by postMessage, so your app draws its own buttons.
Fork a desktop that is signed in. The copy has the same running Chrome and the same session. Or keep the profile in a folder you choose.
Every action through all four. The MCP tools return screenshots as images the model sees.
| What to expect | |
|---|---|
| Network | Open internet by default. Traffic leaves from our servers' addresses, not a home connection: some sites challenge or block sign-ins from data centers. That is what the handoff is for. |
| Proxies | None. We do not route through residential proxies. |
| Passwords and 2FA | Typed by a person in the live view, never by the agent. |
| Downloads | Stay in the VM. |
| Abuse | The same rules and scoring as every sandbox. Input actions are rate limited. |
| Recording | Off unless you ask for it per desktop. |
| What | Time |
|---|---|
| A screenshot, end to end | 365 ms, median |
| A key press to the screen changing, over VNC | 196 ms |
Our own runs against a desktop in production.
| Setting | Value |
|---|---|
| Default size | 2 vCPU, 4 GB |
| Screen | 1280x800 by default, up to 2560x1600 |
| Running | $0.04 per vCPU-hour, like any sandbox |
| Example: 2 vCPU, 4 GB | $0.08 an hour |
| Plan | Paid |