The browser gives an agent a real cloud browser that you can watch and take over.

What it is

The browser gives a harness's agent a real browser in the cloud, driven through typed tools: it opens pages, reads them, clicks, types and takes screenshots. You watch it live in the task, and you can take over at any moment and hand it back.

Turning it on

This replaces the harness's whole plug list, so send every plug it should keep. To give the agent only some of the browser tools, add {"tools": {"browser": ["navigate", "click"]}}, naming the tools without a prefix.

• Connect it once per workspace on the console's Plugins page. It needs no account or key.
• Under Sites, list the only sites it may open, the sites it must never open, or both. A rule covers the domain and its subdomains, and with no list any public site is allowed.
• Then include it on each harness that should browse, in the harness's settings under Plugins. No harness has it until you include it.

POST /v1/harnesses/{id}/servers/plugs
{"plugs": ["browser"]}

What the agent can do

The agent opens a page, reads its address, its text or a snapshot of its elements, clicks, types, presses keys, scrolls, waits for something to appear, goes back, switches tabs and takes screenshots. Screenshots are saved with the task's files.

• It cannot run scripts of its own, download files, read cookies or reach the browser directly.
• Private, loopback and link-local addresses are blocked, and so is any name that resolves to one, for every page and every resource a page loads.
• Harnesses whose agents call tools can use it. System One acts through a fixed set of choices, so browser tools are not offered there.

Watching and taking over

While the agent browses, the task shows the browser live, with a cursor where the agent last acted. Take over, or click the screen, to get the mouse and keyboard: the agent waits, and picks up where it was when you hand the browser back. There is no recording; the screenshots the agent took stay with the task.

GET  /v1/sessions/{id}/browser            open, live_url, control, viewport and how many calls the agent has made
POST /v1/sessions/{id}/browser/control    {"control": "user"} to take over, {"control": "agent"} to hand back

What it costs

Browser time is billed at the provider's price, $0.02 per browser hour, with no markup.

Lifetime and limits

One per task

Opened on the agent's first browser call.

Closed

At the end of each turn, after two idle minutes, or at 20 minutes. A follow-up turn opens a fresh browser.

One call

Up to 60 seconds.

Text read

Up to 20,000 characters per call.

Snapshot

150 elements by default, up to 500.

At once

Up to three open browsers per organization.