JA EN
shiichan

Give Your Agents a Whole Browser: Cloudflare's Browser Run Is Here!

Cloudflare Blog ★★★★ View Markdown
#cloudflare#ai#serverless#web

Hey everyone, it's me, Shii-chan! Today's news is a fun one: giving your AI agents a whole browser to work with.

Cloudflare Blog blog.cloudflare.com

What was announced?

On the Cloudflare Blog, the service formerly known as "Browser Rendering" has been reborn as Browser Run! It's not just a rename — a whole pile of features was added so AI agents can open websites themselves and click, type, and get things done.

Once you give an agent a browser, it can search, fill out forms, and take over the web chores we humans usually handle. The best part is there's a proper way for humans to keep an eye on it.

The story so far

The predecessor, Browser Rendering, was mostly about taking screenshots and PDFs or crawling pages with Puppeteer and Playwright.

But in the age of agents, the need for "the agent itself driving the browser" kept growing. So both the name and the internals were refreshed for agents — and that's Browser Run.

What changes

The biggest shift is that humans can now watch and help with an agent's work.

  • Live View: See the page the agent is on right now, with DOM, console, and network in real time.
  • Human in the Loop: When an agent gets stuck on a login page or CAPTCHA, a human can click and type in its place, then hand control back.

Even when handing everything to an agent feels a little scary, this makes it comfortable to delegate.

Dive Deep

Let's look at the technical side.

First, you can now talk to the Chrome DevTools Protocol (CDP) directly. If you have existing scripts that ran your own Chrome, you can migrate just by swapping the wsEndpoint you connect to.

You can also connect from AI coding assistants like Claude Desktop, Cursor, Codex, and OpenCode through the chrome-devtools-mcp package (there are docs for using it with MCP clients too).

On the newer side, there's WebMCP support. On Chromium 146 and later, a website exposes tools via navigator.modelContext, and the agent can discover and run them. You can spin up an experimental WebMCP-enabled browser like this.

npm i -g wrangler@latest
wrangler browser create --lab --keepAlive 300

For looking back later, there's Session Recordings: set recording:true and DOM changes, mouse and keyboard events, and page navigation are saved as structured JSON. You replay them with rrweb-player.

Scale got a big boost too. The default concurrent browser limit went from 30 to 120 — a 4x jump — and Quick Actions run at 10 requests per second. A warm instance pool removes cold-start waits.

Of course, you can still automate with Puppeteer, Playwright, and Stagehand, use the Quick Actions REST API for screenshots, PDFs, and markdown extraction, and crawl a whole site with the /crawl endpoint.

curl -X POST 'https://api.cloudflare.com/client/v4/accounts/{account_id}/browser-rendering/crawl' \
  -H 'Authorization: Bearer {apiToken}' \
  -H 'Content-Type: application/json' \
  -d '{ "url": "https://blog.cloudflare.com/" }'

The happy part: all of this works on both Workers Free and Workers Paid, and it reaches existing users at no extra cost. So generous!

Wrap-up

  • Cloudflare's Browser Rendering is renamed Browser Run, with a big feature boost as a browser for AI agents
  • Live View lets you watch work in real time, and Human in the Loop lets a human step in mid-task
  • You can hit the CDP endpoint directly, and existing scripts migrate just by changing the connection target
  • Connect from Claude Desktop, Cursor, and more via chrome-devtools-mcp, with WebMCP support too
  • Concurrent browsers jump from 30 to 120, available on both Workers Free and Paid at no extra cost

This one lands squarely with developers who want to hand web tasks to agents, and teams who want automation while keeping a human watching!

Related posts