Skip to the yard
Waxsealai

Cloudflare Opens Kitesurf, Its Workers-Based Browser for AI Agents

Shipped
Shipper
Celso Martinho
Readers
686

Kitesurf’s Apache 2.0 release lets developers inspect, modify and deploy the agent-focused browser, with compatibility and reliability still under active development.

Cloudflare Opens Kitesurf, Its Workers-Based Browser for AI Agents

Kitesurf is now available under the Apache 2.0 license, giving developers access to a browser designed for AI agents and built to run on Cloudflare Workers. They can inspect and change its code, run it locally, or deploy a separate instance to their Cloudflare account. The release makes the implementation available; it doesn’t mean the browser has finished closing the compatibility gap with established browsers.

The project was introduced in August and updated in September with WebMCP support, additional APIs and improved compatibility. Its central idea is to handle browser tasks without relying on a conventional browser container. Kitesurf parses HTML, applies CSS, runs page JavaScript and produces screenshots, while exposing the Chrome DevTools Protocol (CDP) so automation clients such as Puppeteer can connect.

A different browser architecture

Kitesurf’s design targets workloads where an agent needs to load a page, inspect or interact with it, and capture the result. The project aims to use less CPU and memory than a full browser such as Chromium, while running across Cloudflare’s Workers network and scaling on demand. Those are design goals; the project points developers to its compatibility tracking and benchmarks for evidence as the engine develops.

The implementation is divided into three main parts. The Engine handles browser sessions and client connections. PageScript runs page code, including JavaScript. PageRenderer draws the page’s output. Together, these components provide a browser interface without making a conventional browser container the foundation of each session.

That architecture brings practical tradeoffs. A browser engine has to interpret web markup, styles and scripts in ways that many sites expect, and small differences can affect rendering or automation. Kitesurf’s development priorities include improving compatibility and reliability, so teams evaluating it should check their own pages and workflows rather than assume every site behaves as it would in Chromium.

Rust and WebAssembly components

The repository combines Cloudflare’s developer platform and APIs with Rust projects compiled to WebAssembly. Blitz handles HTML parsing, the document object model (DOM), layout and painting. Stylo handles CSS parsing, cascading and computed styles. For JavaScript, Kitesurf can use Dynamic Workers; Boa runs in WebAssembly as a fallback when JavaScript can’t run in Dynamic Workers.

Other dependencies cover text layout, graphics and document output. Linebender libraries support text layout and two dimensional graphics, while Krilla generates PDF output. These projects keep their own licenses. The Kitesurf repository’s third party notices documents those dependencies, so teams assessing reuse should review the relevant licenses alongside Kitesurf’s Apache 2.0 terms.

The project’s packages, Rust core, DOM shims, renderer and command line interface are organized separately in the repository’s packages directory. That layout gives developers a way to follow how the browser’s parts fit together and focus changes on a specific component.

Local development and automation

To start a local instance, developers first need the build requirements, then can clone the repository and run:

git clone https://github.com/cloudflare/kitesurf.git
cd kitesurf
pnpm install --frozen-lockfile
pnpm run dev

The initial build compiles the Rust core and its dependencies, which takes longer than later builds. Once compilation completes, Wrangler starts a development server at http://127.0.0.1:8787. Developers can open that address in Chrome DevTools or connect automation software to the local browser interface.

For example, Puppeteer can connect over WebSocket using Puppeteer’s connection interface:

import puppeteer from "puppeteer-core";
const browser = await puppeteer.connect({
  browserWSEndpoint: "ws://127.0.0.1:8787/devtools/browser",
});
try {
  const page = await browser.newPage();
  await page.goto("https://example.com/");
  await page.screenshot({ path: "page.png" });
} finally {
  await browser.disconnect();
}

Agent frameworks can also connect through CDP and control the browser. This gives developers a familiar automation path for trying Kitesurf with existing tooling, though a successful connection alone doesn’t establish that every browser feature or page will work as expected.

Deploying and assessing the project

After local development, the engine can be deployed to a developer’s own Cloudflare account with the repository’s deployment steps. The deployment command is:

pnpm run deploy

A Workers Paid plan is required because Kitesurf uses Dynamic Workers to create isolated Workers for pages. That account and plan requirement is a concrete consideration for teams comparing a self managed deployment with the Browser Run managed APIs.

Kitesurf remains under active development, with compatibility, reliability and efficiency still on the project’s worklist. Developers can follow changes in Web Platform Test results and the project’s benchmarks, then test the specific sites and tasks their agents depend on. The repository welcomes contributions, including focused code changes, reproductions of rendering bugs, automation failures and documentation fixes. Its contribution guidelines ask contributors to open an issue with a minimal reproduction or proposed approach and receive maintainer support before submitting a pull request.

The release gives engineering teams a way to examine and run the browser themselves, while leaving an important evaluation task in their hands: measure how its current rendering and automation behavior holds up on their workloads, and track whether compatibility progress meets their needs.

Source: Celso Martinho · blog.cloudflare.com

Same vesselRelated cargo