Kitesurf: Cloudflare's Browser Built for AI Agents

Kitesurf: Cloudflare's Browser Built for AI Agents

Kitesurf is Cloudflare's browser for AI agents: 3 to 7 times less CPU and memory than Chromium, but slower and not able to run everything. Here is what to check on your site.

Kitesurf is a browser Cloudflare built for AI agents. It runs entirely on Cloudflare Workers and is available now, free during the beta, inside Browser Run. According to Cloudflare, it uses 3.1 to 7 times less CPU and memory than Chromium (the engine behind Chrome) on common agent tasks such as taking screenshots and extracting HTML. The price is speed: each task takes about 1.7 to 1.8 times longer.

For anyone running a website, the consequence is practical. The cheaper it gets to give every agent a browser, the more agents will open your site, and they will see the page differently from the Chrome you use for testing. The announcement is on the Cloudflare blog, signed by Celso Martinho, Ruskin Constant, Rui Figueira and Luís Duarte.

What is Kitesurf and why did Cloudflare build a browser?

Kitesurf is a browser that drops everything only humans need. Cloudflare says engines like Chromium were built for people and carry a memory and compute cost an AI model does not need. An agent does not care about tabs, themes, extensions or syncing across devices. It cares about token count, context window, scale and cost.

The company also argues that Chromium is too expensive to give each agent its own instance, which would leave much of the web reserved for the largest and priciest models. Kitesurf was built to change that math. The project is 12 weeks old, with its first commit in May, and Cloudflare promises to open-source it once it is ready.

It speaks CDP (Chrome DevTools Protocol), so Puppeteer, Playwright and MCP-based agents work unchanged. You only add browser=kitesurf to the Browser Run endpoint.

How much does Kitesurf save compared to Chromium?

Kitesurf saves CPU and memory and loses on wall-clock time. Cloudflare measured the medians of five Browser Run quick-action runs over a 14-URL corpus, against Chromium with a warm pool:

Metric Kitesurf Chromium (warm pool) Kitesurf, relative
CPU: screenshot 380 ms 1,173 ms 3.1x less
CPU: HTML extraction 229 ms 877 ms 3.8x less
Memory: screenshot 57.8 MiB 271.0 MiB 4.7x less
Memory: HTML extraction 39.4 MiB 273.7 MiB 7.0x less
Wall time: screenshot 1,148 ms 637 ms 1.8x slower
Wall time: HTML extraction 820 ms 472 ms 1.7x slower

These are Cloudflare’s own numbers, from a small corpus, and have not been independently tested. Its explanation for the delay is that a JIT that has already seen a page always beats a cold software renderer. Much of the gap comes from rasterization and JPEG/PNG encoding.

What can’t Kitesurf do?

Kitesurf cannot yet play video, render WebGL, negotiate a bot-challenge handshake with real TLS fingerprints, or hold a ten-minute authenticated session that needs persistent state. In those cases Cloudflare says to use the Browser Run default, which is still Chromium.

Today it renders TodoMVC (vanilla, React, Vue, Angular and Preact), Wikipedia, Hacker News, the Cloudflare Blog and much of the Cloudflare dashboard. The company says it passes more than 215,000 tests from the Web Platform Tests, the public conformance suite for web standards, and adds hundreds more every week.

One technical detail that matters for your site: JavaScript runs, but eval is not natively supported in Workers. Cloudflare works around it with Boa JS, an interpreter written in Rust, and admits it is “a runtime on top of a runtime”. It handles the occasional eval, and nothing more.

How does Kitesurf change what an agent sees on your site?

Kitesurf changes what the agent sees because its goal is structured content, not visual fidelity. Cloudflare writes that an agent is fine with slightly off CSS or rendering that is not pixel perfect. What it wants is the DOM and the text.

When I got into SEO in 2014, the question was whether Googlebot executed your JavaScript. The question now is similar, with more creatures at the table: can a user’s agent, running on a cheap browser, reach the content you want it to cite? Cloudflare’s post does not say how Kitesurf identifies itself or how Bot Management treats it, so you will have to check that in your own logs.

If you block bots, it is worth rereading what I wrote on blocking AI training without disappearing from Google and on how most of the crawler traffic in your log is fake. A new agent browser is one more entry in that log you need to know how to read.

What should you do on your site on Monday?

The first step is to test your site in the public Kitesurf playground, where you can type any URL, see how the page comes out and open the embedded DevTools. Cloudflare recommends exactly that: the best way to know whether a site is compatible is to try it.

Then, a short check:

  1. Open an important page (product, service, article) and confirm that the main text, price and contact details are in the HTML or appear without depending on WebGL, video or interaction.
  2. Compare the result with your own Chrome. If Kitesurf shows an empty page where you see content, the low-cost agent will see it empty too.
  3. Watch your server or CDN logs over the next weeks for new access patterns, without assuming a user-agent name: Cloudflare did not disclose one.
  4. If critical content only loads after an authenticated session or a bot challenge, a lightweight agent will probably be left out. Decide whether that is what you want.

What does Kitesurf mean, in short, for site owners?

Kitesurf is a beta browser that makes AI agent access to the web 3 to 7 times cheaper in CPU and memory, at the cost of being about 1.7 to 1.8 times slower and not running video, WebGL, bot challenges with real TLS or long stateful sessions. Cloudflare has not yet open-sourced it or said how it identifies itself.

Your side of the decision is simple: make sure the content that matters reaches a lean browser as renderable HTML, test it in the playground and keep an eye on the logs. If your business depends on being cited by AI assistants, this short test belongs in your routine next to checking what Google treats as the main content of a page.

Back to Blog