Cloudflare Built a Browser for AI Agents, and Chromium Is Still Faster
On 6 August 2026, Cloudflare announced Kitesurf, a headless browser written for AI agents rather than for people, and running entirely inside V8 isolates on Cloudflare Workers. The interesting part of the announcement is not the one being repeated everywhere. Kitesurf consumes far less CPU and memory than Chromium — but on the published benchmarks it is also slower in wall time. That combination is the whole story, and it says something precise about what agent infrastructure is actually optimising for.
What it is, technically
Kitesurf is not a Chromium fork. It is assembled from Rust components compiled to WebAssembly: Blitz handles HTML and layout, Stylo — Firefox's CSS engine — parses the stylesheets, and Boa covers JavaScript evaluation. The whole thing runs in the same V8 isolates that already power Workers, which is what makes the memory numbers possible: there is no separate browser process to start, sandbox and tear down for each page.
It speaks the Chrome DevTools Protocol, so the existing tooling works unchanged: Puppeteer, Playwright, chrome-remote-interface, and MCP-compatible agents all connect to it the way they connect to Chromium. Enabling it is a query parameter — browser=kitesurf on a Browser Run endpoint. Cloudflare reports that it passes more than 215,000 Web Platform Tests, with coverage across CSS, DOM, HTML, selection, SVG and XHR.
The numbers, including the ones that don't flatter it
Cloudflare published median results against Chromium on two workloads. Taking a screenshot: 380 ms of CPU against 1,173 ms, and 57.8 MiB of memory against 271.0 MiB. Extracting HTML: 229 ms of CPU against 877 ms, and 39.4 MiB against 273.7 MiB. That is roughly 3 to 4 times less CPU and 5 to 7 times less memory.
And then the figure that most coverage leaves out: Chromium finishes first — about 1.8 times faster on screenshots and 1.7 times faster on HTML extraction. Kitesurf is not a faster browser. It is a cheaper one per unit of work, which is a different claim and, for the workload it targets, a more useful one. An agent fleet does not care whether one page took 400 ms or 700 ms; it cares how many pages it can hold open at once before the bill or the memory ceiling stops it.
What it deliberately cannot do
The list of omissions is as informative as the feature list. Kitesurf does not play video and does not render WebGL. It does not handle TLS fingerprinting challenges, and it is not built to hold a long authenticated session open. It does not aim for pixel-perfect rendering. Every one of those is something a human user needs and an extraction agent does not — which is precisely the bet: an agent wants machine-readable content, low token overhead, and isolation, not a faithful reproduction of what a person would see.
That bet has a practical edge. If your agent's job is to open a page, read the DOM and hand structured content back to a model, then video decoding and WebGL are pure cost. If your job is to test a real user journey through an authenticated dashboard, Kitesurf is the wrong tool today and Cloudflare says so plainly.
Availability, and what remains open
Kitesurf ships inside Browser Run and is free while in beta, subject to per-account limits. The team states it intends to open source the project "once we're ready". Until that happens, the benchmarks above are Cloudflare's own, run on Cloudflare's platform — worth keeping in mind before treating the ratios as universal. We have not re-run them independently.
Why this matters beyond Cloudflare
For twenty-five years, every headless browser has been a human browser with the screen removed. Kitesurf is the first widely available one designed the other way round — from the agent's requirements inward. It arrives alongside the standardisation of MCP as the agent tooling protocol, and it runs on the same edge execution model that already reshaped how we deploy web applications.
The practical question for anyone building agents right now is narrower than the headlines suggest: are you paying for pixels you never look at? If a meaningful share of your browser bill goes to rendering fidelity that only a human would notice, an agent-first engine changes your unit economics. If your agents genuinely need to behave like users — logging in, holding sessions, dealing with anti-bot challenges — then Chromium remains the answer, and will for a while. The honest reading of this launch is not that Chromium has been beaten. It is that the browser has finally stopped being assumed to have a human on the other side.
