Why give the AI a browser
Building a website from a prompt only gets you so far when the target is "match this competitor's homepage" or "check what my published site actually looks like right now." Until now, the AI had no way to look at a real, live webpage — it could only reason about text you pasted in. That gap is closed.
The AI can now navigate to any real URL, inspect what's actually rendered — headings, buttons, links, colors — and act on what it finds, the same way a person would: click something, scroll down, look again.
Browser agent in chat
In the AI chat panel, admins can toggle on the Browser agent (the globe icon next to the message input). Once enabled, describing a task like "go check what colors dartanger.com uses on their homepage" hands control to the browser agent — it opens the page, reads it, and reports back.
We also added a separate Search icon right next to the attach button that just toggles the preview panel over to the live browser view — no agent activation, no side effects on the model or chat routing. It's a pure "let me look at what's in the browser panel right now" toggle, completely independent of the agent itself.
Browser tool in the Pipeline Builder
The bigger addition is in the Pipeline Builder: a new Browser capability, enabled the same way as Read, Gather, Database, or Edit on any agent step. Once enabled, the AI can write a self-closing tag to control the browser mid-pipeline:
<browser action="open" url="https://example.com" />
<browser action="get_page_data" />
<browser action="scroll" percent="100" />
<browser action="click" ref="e3" />
<browser action="screenshot" />This means a pipeline step can genuinely research a live site — open it, read what's there, scroll to see more, click through to another page, and capture a screenshot — all as part of an automated, repeatable workflow, without any manual steps.
What the AI can actually do
- Open — navigate to any URL, in chat or in a pipeline step
- Read the page — get section headings, every clickable button and link (each with a stable reference id), and the page's dominant colors
- Scroll by percentage —
percent="100"scrolls straight to the bottom of the page, useful for long landing pages or infinite-scroll feeds - Click — using the reference id from the page read, so the AI clicks the actual button or link it just identified, not a guess
- Screenshot — captured and handed back to the AI so it can visually confirm what it's looking at, not just infer from text
Watching it happen live
The preview panel automatically switches to show the live browser the instant the AI starts using it — in chat or mid-pipeline — and switches back to your normal app preview the moment the AI moves on to a different kind of step. You never have to manually flip a toggle to follow along; it tracks what the AI is actually doing in real time.
This matters most in pipelines: if a step reads a competitor site, then a later step edits your own code, the preview will show you the browser during the research step and your app's preview once the AI starts writing.
Under the hood, this runs on a real headless browser driven over the Chrome DevTools Protocol — the same low-level interface browser automation and testing tools use — so what the AI clicks and reads is exactly what a real visitor would see.
Everything at a glance
| Capability | Where | What it does |
|---|---|---|
| Browser agent toggle | Chat panel (admin) | Hands the conversation to an AI-driven browser session |
| Search icon | Chat panel (everyone) | Just shows the browser view in preview — no agent side effects |
| Browser capability | Pipeline Builder | Lets any agent step open, read, scroll, click, and screenshot real pages |
| Live preview switch | Both | Preview panel follows what the AI is doing automatically |
Both entry points are live now. Try it from the chat panel's globe icon, or add the Browser capability to a Pipeline Builder step.
Related: How to automate your dev workflow with the Pipeline Builder · Total control over your AI agent