🎧 Prefer to listen?

If you’ve ever tried to build an AI agent that browses the web, you’ve hit the wall: running a full browser for every agent session is expensive, slow, and overkill for what the agent actually needs. Cloudflare just built something that might change the equation — a browser engine from scratch that’s designed for machines, not people.

The problem nobody talks about

Every AI agent that needs to interact with the web — whether it’s scraping data, filling out forms, or taking screenshots — runs into the same bottleneck. You need a browser. And the only browsers available were built for humans.

Chromium, the engine behind Chrome, Edge, and most headless browser tools, is a remarkable piece of software. It’s also absurdly heavy for what most agents actually do. A single headless Chromium instance eats over 250MB of RAM and burns through more than a second of CPU time just to render a page and take a screenshot. Multiply that by however many concurrent agents you’re running, and you’re either paying for a fleet of beefy VMs or rationing browser access so heavily that only the most expensive AI models get to use the web.

I ran into this myself when testing automated web scraping workflows — the browser was always the most expensive part of the pipeline. Not the AI model, not the data processing. The browser.

What Cloudflare actually built

Kitesurf is Cloudflare’s answer to this problem. It’s not a wrapper around Chromium. It’s not a fork of Firefox. It’s an entirely new browser engine built from scratch in Rust and WebAssembly, designed to run inside Cloudflare Workers — their serverless platform that spins up isolated V8 environments on demand.

The key insight is simple: most AI agent tasks don’t need a full browser. They need to parse the DOM, run JavaScript, extract text or take a screenshot, and move on. They don’t need tabs, bookmarks, extensions, smooth scrolling, or pixel-perfect CSS rendering.

So Cloudflare stripped all of that out. Kitesurf has four components, each running as a separate stateless Worker:

  • Engine — speaks the Chrome DevTools Protocol (CDP), so existing tools like Puppeteer and Playwright connect to it without code changes
  • PageScript — parses HTML and CSS using Blitz and Stylo (Firefox’s CSS engine), runs JavaScript with Boa (a Rust JS interpreter)
  • PageRenderer — rasterizes pages into screenshots or PDFs
  • SandboxOutbound — the only component with network access, enforcing security boundaries

The stateless design is the real trick. Each page load spins up an isolate, does its work, and tears down. No warm pool of browser processes eating resources between requests. For agent workloads — which are bursty, unpredictable, and often abandoned mid-session — that’s exactly the cost profile you want.

The numbers

Cloudflare benchmarked Kitesurf against Chromium across 14 URLs, run five times each, on the two tasks agents run most: screenshots and HTML extraction.

For screenshots: Kitesurf uses 3.1x less CPU and 4.7x less memory than Chromium. For HTML extraction: 3.8x less CPU and 7x less memory.

Those are real savings that directly map to your cloud bill. But there’s an honest tradeoff — Chromium is still 1.7x to 1.8x faster in raw wall-clock time because of its JIT compilation. If your agent needs the absolute fastest response on a single page, Chromium wins. If you’re running thousands of concurrent agent sessions and cost matters more than latency, Kitesurf wins by a mile.

What this means if you’re building with AI agents

If you’re a solo builder using AI agents for web automation, data extraction, or agentic workflows, here’s what changes:

Cost drops dramatically. Running a browser per agent session was the hidden cost that made many agent architectures impractical. Kitesurf makes it viable to give every agent its own browser session without provisioning expensive VMs.

Scaling gets simpler. Because Kitesurf runs as stateless Workers, you don’t manage a pool of browser instances. Cloudflare handles the scaling. You pay for what you use, not for idle capacity.

Existing tools still work. Kitesurf speaks CDP, so your Puppeteer or Playwright scripts connect to it the same way they’d connect to Chromium. You add browser=kitesurf to the endpoint and nothing else changes. That’s the kind of backward compatibility that makes adoption realistic.

The web becomes more accessible to cheaper models. This is the bigger picture. When browser access is expensive, only the most sophisticated (and costly) AI models can afford to browse the web. Kitesurf lowers that barrier, which means smaller models and budget-constrained projects can compete on tasks that previously required premium infrastructure.

What it can’t do yet

Kitesurf is in beta, and there are real limitations:

  • No video playback or WebGL
  • No TLS fingerprinting, which means some bot-detection challenges will block it
  • No persistent sessions — it can’t maintain long-lived authenticated logins
  • It runs on Cloudflare’s network only — it can’t access localhost on your machine without a tunnel
  • CSS rendering is good enough for extraction but not pixel-perfect

If your agent needs to watch a video, play a game, or maintain a logged-in session across multiple steps, Chromium is still your tool. But for the majority of agent tasks — “go to this page, extract this data, take this screenshot” — Kitesurf is purpose-built.

How to try it

Kitesurf is available now in Browser Run, Cloudflare’s existing headless browser API. If you’re already using Browser Run with Chromium, switching is literally one parameter change. If you’re not, Browser Run has a free tier that’s enough to test with.

The broader context here is Cloudflare’s push toward what they call the “Agentic Internet” — infrastructure designed for AI agents as a first-class user, not an afterthought. Kitesurf is one piece of that puzzle, alongside their work on AI search, agent behavior detection, and stablecoin payments for agents.

The bottom line

Kitesurf won’t replace Chromium for human browsing, and it’s not trying to. But for the growing number of solo builders running AI agents that need to interact with the web, it’s the first browser that was actually built for the job. The cost savings alone make it worth testing — and the fact that it works with your existing tools means there’s no migration cost to find out if it fits your workflow.

If you’re building anything that involves AI agents and the web, start here: Cloudflare Browser Run.

Want to see what else is possible with AI automation? Check out the tools I actually use every day or explore how to build your first automation in 15 minutes.