Kitesurf, Cloudflare's agent browser, now lets a website expose functions an agent can call instead of clicks it has to simulate
Cloudflare's Kitesurf agent browser has added WebMCP support so a website can hand an agent a function like searchFlights() instead of asking it to guess where to click, and it now passes 730,000 or more Web Platform Tests.

Image: Cloudflare
Why it mattersA team building an agent that shops, books or files things through websites can now target a stable function interface where the site publishes one, which is the difference between a script that works today and one that works next quarter.
An agent asked to book a flight has to open a browser, find the search box, guess the layout, type into it and hope the site did not move a button that morning. That is the whole reason agent browsing breaks so often. Cloudflare's answer is to skip the clicks when the site offers to.
Cloudflare shipped an update to Kitesurf on 28 September 2026. Kitesurf is a browser built on Cloudflare Workers for AI agents; it launched in beta in August 2026 and this update adds WebMCP, changes how the browser runs JavaScript, and lets you point it at a terminal.
Calling a function instead of a click
WebMCP is the piece that matters most. It lets a website publish functions in a machine-readable form: searchFlights(from, to, date), addToCart(sku), filePermit(fields). When a Kitesurf agent lands on a page that offers WebMCP, it can call those functions directly through the Model Context Protocol rather than driving the DOM. Cloudflare's own words: agents "no longer need to simulate clicks and typing".
That does not remove the fallback. A page without WebMCP is still a page an agent can drive with the Chrome DevTools Protocol, Playwright or Puppeteer, which Kitesurf continues to support through its Browser Run API. WebMCP is the fast path when a site publishes one, and Cloudflare adds a env.BROWSER.quickAction() call in Worker scripts to trigger it.
The Web Platform Tests number
Kitesurf now passes "over 730,000 subtests" from the Web Platform Tests, up from around 230,000 at launch. Web Platform Tests are the cross-browser conformance suite that Chrome, Firefox and Safari run against, and a higher number means a wider slice of the real web renders correctly. That is Cloudflare's own count against a public test suite, not an independent audit.
Cloudflare says it got there through work at the boundary between the Boa JavaScript engine and the DOM: fewer crossings between the two, lighter timers and script loading, and font fetching that pulls only the character sets the page uses.
Running the browser in a terminal
Kitesurf can now render into a terminal, either as ANSI text or as inline images through the Kitty graphics protocol. Cloudflare positions this at the developer building an agent: seeing what the agent is looking at, without a graphical desktop, is often the whole difference between a bug that reproduces and one that does not.
The limits, said plainly
Kitesurf is in free beta with per-account limits, and Cloudflare says the open-sourcing of the browser code is "coming soon" but has not happened yet. So the developer story today is: use it through the Browser Run API, and treat the beta caps as the thing you have to plan around. The WebMCP promise is only as good as the sites that publish it: on a page without WebMCP, an agent still clicks and types, and every fragile assumption that has ever broken a browser bot still applies.
The direction is more useful than the release. A browser designed for agents rather than for people is a different bargain: fewer human-facing niceties, and a formal interface where one exists. If enough sites decide to publish a WebMCP surface next to the human one, the agents that book their flights and file their forms will stop breaking every time the marketing team redesigns the page.
Source
- Cloudflare blog: The road to the agentic browser: A Kitesurf update
- Browser Run API docs: developers.cloudflare.com/browser-run
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


