Cloud Computer

Click Deploy on the Cloud Computer card. You get COMPUTERS_API_URL and COMPUTERS_API_KEY. From your server, ask for a computer for one of your users, show its screen in your app, and let your AI agent use it: take a screenshot, send it to a model, then click, type or run a command with what the model answers.

• Each computer is a Linux desktop (Debian with Xfce) with Google Chrome as its default browser, Chromium, a terminal and a file system of its own, running in an isolated sandbox.
• Give each of your users their own computer by passing your id for them as userId. A computer keeps its user's files, browser sign-ins and screen, so never share one between different people.
• Asking for a userId that already has a computer starts that computer and returns it, so your app does not need to store computer ids.
• Try computers from the dashboard too: Cloud Computer → New computer opens its live screen, and you can use it with your mouse and keyboard or try agent actions.
• Keep the key on your server. It is not meant for browser code.

const api = (path: string, body?: unknown) =>
  fetch(`${process.env.COMPUTERS_API_URL}${path}`, {
    method: body === undefined ? "GET" : "POST",
    headers: {
      authorization: `Bearer ${process.env.COMPUTERS_API_KEY}`,
      "content-type": "application/json",
    },
    body: body === undefined ? undefined : JSON.stringify(body),
  }).then((r) => r.json());

// Creates and starts this user's computer, or starts the one they have.
const computer = await api("/computers", { userId: user.id, size: "basic" });

// A link to the live desktop, for an <iframe> in your app. It works for an hour.
const { url } = await api(`/computers/${computer.id}/preview`, { viewOnly: false });

Pass viewOnly: true to let your user watch the agent without taking over the mouse and keyboard. Ask for a new link when the old one expires. When your user leaves, POST /computers/:id/disconnect ends every live view at once, so a link they kept stops working.

Each step of a computer-use agent is one POST /computers/:id/actions request. Screenshots come back as a base64 PNG to send to your model; positions are pixels from the top left of the screen (1280 × 800).

const act = (action: object) => api(`/computers/${computer.id}/actions`, action);

const shot = await act({ type: "screenshot" }); // { image, mime, cursor }
await act({ type: "click", x: 640, y: 400 }); // button: "left" | "right" | "middle", double: true
await act({ type: "move", x: 100, y: 200 });
await act({ type: "drag", x: 100, y: 200, toX: 400, toY: 200 });
await act({ type: "scroll", x: 640, y: 400, direction: "down", amount: 3 });
await act({ type: "type", text: "hello world" });
await act({ type: "key", key: "enter", modifiers: [] });
await act({ type: "hotkey", keys: "ctrl+l" });
const out = await act({ type: "exec", command: "ls ~", timeout: 30 }); // { exitCode, output }

Anthropic's and OpenAI's computer-use tools both work this way: turn each tool call the model makes into one of these actions, and send back the next screenshot.

• Send the key as Authorization: Bearer <key>. Errors are { "error": { "message" } } with a 4xx or 5xx status.
• Starting is refused with 402 when the account is out of credits, at its spend cap or past its monthly computer allowance, and with 403 when the project keeps as many computers as the plan allows.

Only running time is charged, counted by the second, from your credits. A stopped computer keeps its files at no charge. Usage is charged every hour and shows as Cloud Computer on Credits & Billing.

• A computer with no mouse, keyboard or API activity stops by itself. Change how long it waits under Cloud Computer → Settings, from 1 minute up to your plan's longest; running computers take the new time at once.
• Stopped for a day, a computer is archived: its files are kept and the next start takes a little longer.
• Running computers are stopped when the account runs out of credits or reaches its monthly computer allowance.