Cloudflare Kitesurf uses 3 to 7 times less CPU and memory than Chromium for the tasks AI agents actually do.
And it is free in beta through Browser Run.
It is a stateless, agent-first browser built from scratch in Rust compiled to WebAssembly, running inside V8 isolates on Cloudflare Workers. Rather than hauling a full desktop engine around, Kitesurf optimizes for what a model cares about and drops tabs, themes, extensions, and pixel-perfect rendering. Cloudflare describes it as “stateless, highly scalable” and “designed for AI agents.” If you run browser-based agent workflows today, this is the first major engine built for the thing on the other end of the connection: not a human, a model.
What Cloudflare Kitesurf Actually Is
Kitesurf runs entirely on top of Workers, Cloudflare’s serverless platform.
And it was designed specifically for what Cloudflare calls the Agentic Cloud. The pitch is worth taking literally: AI developers can now build software that navigates websites, fills out forms. And completes browser-based tasks without shipping their own browser software. That last line matters more than it sounds. Every headless-Chromium setup I have stood up for a client eventually becomes a maintenance project, with version pinning, font installs, sandbox escaping, and zombie processes.
A stateless browser that spins up in a V8 isolate and disappears when the job is done removes a whole class of that pain.
The “stateless” framing is the part most people will gloss over and the part that actually changes the architecture. Stateless means no persistent profile, no session cruft, no tab state leaking between runs. For an agent doing one-shot page reads, that is exactly what you want. For anything that needs login state or a warm cache across steps, it is a constraint you have to design around rather than ignore.
Why A Browser For Agents Makes Sense
Cloudflare’s own framing is the clearest version of the argument I have read: “We should be giving all agents tools that excel at what’s important for an AI model.” The standard agent stack runs headless Chromium, which was built for a human sitting at a desk. That engine assumes you care about tabs, themes, extensions, pixel-perfect rendering, and 60-fps scrolling. None of that helps an agent that just needs a screenshot and the HTML.
Kitesurf focuses on what matters to an agent instead: token count, context windows, scalability, performance, and cost. Every one of those is a real constraint in production. Token count is what you pay for on every page you feed a model. Context windows cap how much of a page you can reason over. Cost is the line item that decides whether a client workflow is viable or gets cut. A browser that treats those as the primary metrics is a browser built by people who have actually run agents, not just spec’d them.
Rust, WebAssembly, And A 12-Week Build
The engineering story is its own signal. Cloudflare wrote a browser engine in Rust that runs inside Workers V8 isolates, compiled to WebAssembly. The project went from first commit to public beta in 12 weeks. That is not a browser-engine timeline by any normal measure. It is a timeline you can hit only when you are willing to delete most of what a browser does.
The Rust-to-WASM-to-V8-isolate path is also why the resource numbers land where they do. V8 isolates are cheap to spin up and tear down, which is the whole point of running this on Workers rather than on a fleet of containerized Chrome instances. daily.dev reports Kitesurf already passes 215,000-plus Web Platform Tests, which is the credibility claim that separates a real browser from a toy renderer.
Passing that many conformance tests in 12 weeks tells you the team reused proven components rather than rewriting CSS from zero.
The Benchmarks And The Catch
Here is where it gets honest. On a 14-URL test corpus, Cloudflare reports Kitesurf used 3.1x less CPU for screenshots and 3.8x less for HTML extraction compared to Chromium. Memory was 4.7x lower for screenshots and 7x lower for HTML extraction. Those are the headline numbers. And they are vendor-reported on a curated corpus, so read them as a best case.
The catch Cloudflare did not bury: Kitesurf is 1.7x to 1.8x slower in wall time because there is no JIT. A cold software renderer cannot beat a warm just-in-time compiler, and they said so plainly. For my work that tradeoff is usually fine. Most agent browser jobs are not latency sensitive; they are throughput and cost sensitive. I would rather take the memory savings and run more concurrent sessions than chase a faster wall time on a single page grab. But if you have a workflow where a human is waiting on the result of each step, that 1.7x penalty compounds across a multi-step chain and you will feel it.
The part nobody has published yet is an actual dollar comparison.
Cloudflare says this lowers the cloud bill.
And the resource math makes that plausible.
But there is no public pricing breakdown of Kitesurf sessions versus Chromium sessions on Workers. If you are evaluating it for a paying client, run both on the same workload and measure your real bill. The benchmark corpus is 14 URLs.
And nobody has shown how Kitesurf behaves on a JavaScript-heavy enterprise SPA or a strict anti-bot wall, which is exactly where agent stacks break today.
What This Means For Small Operators
For a solo or small agency, the beta being free is the on-ramp. Kitesurf is available for free while in beta in Browser Run. And existing Puppeteer, Playwright, and MCP clients connect via CDP with no rewrite. That last detail is the real unlock. You do not have to throw away your automation code to try it. You point your existing client at Kitesurf and measure the difference on a real job.
The bigger tension is worth naming. The same enterprise that runs much of the web’s bot blocking and DDoS mitigation now too ships the browser that agents browse with. That is not a reason to avoid Kitesurf. It is a reason to read the isolation and security model carefully before you route sensitive workflows through it. And to keep a Chromium fallback for anything the beta cannot handle yet.
The wall-time penalty alone justifies a hybrid setup for time-sensitive steps.
Here is my read. Kitesurf is the first credible sign that “the browser” is splitting into two products: one for humans and one for agents. Chromium will not disappear, but the idea that every agent stack should inherit a human’s browser is already obsolete. The teams that win on agent economics will be the ones who stop paying for 60-fps scrolling nobody watches.
If you run browser-based agents, point your real workload at the Browser Run beta this week. Compare the resource numbers and the wall time against your Chromium setup on the same job, not on a vendor’s test corpus.
That comparison is the only data that matters for your bill.
