Cloudflare open-sources Kitesurf, a browser engine without Chrome

An AI agent does not necessarily need a full Chrome installation to use a website. Cloudflare announced on November 12 that Kitesurf, its experimental browser engine, is open source. It can process a page, execute its JavaScript and produce screenshots without launching Chrome or a browser container. Developers can inspect and change the code and operate an instance in their own account. That turns the engine itself into something teams can adapt, rather than only a remote service they can call.

Invisible browsers are not new. Google’s documentation describes Chrome Headless as a way to run Chrome without a visible interface in an unattended environment. Modern Headless shares its implementation with ordinary Chrome. Hiding the window therefore does not remove the machinery behind it. Kitesurf takes a different architectural route. For a team choosing tools, retaining a complete browser and assembling an engine around automated tasks are different approaches, with different compatibility questions to answer.

The project first appeared in August. By September, it supported WebMCP, a mechanism for websites to offer agents callable tools. An agent might invoke a flight-search function instead of working out which pixels to click. In its September 28 update, Cloudflare also reported passing more than 730,000 web-platform subtests, roughly 500,000 more than at launch. Those are individual tests, not a count of websites that work. A page combines many behaviors, and getting its text and buttons onto a screen is only part of matching a mature browser.

Kitesurf divides its main work among three components: one handles sessions and client connections, another runs the page and its scripts, and a third draws the result. Think of a small restaurant separating orders, cooking and plating. Cloudflare’s November 13 technical explanation also credits existing open-source components for layout, styling and graphics. The company’s aim is lower resource use than a full browser; compatibility and performance remain active work. The novelty is a different assembly of browser capabilities, not the claim that automated browsing has just been invented.

Operating it brings concrete responsibilities. The repository says a Cloudflare deployment requires a paid Workers plan, and the engine has no authentication of its own. Operators need a separate access-control layer, or someone else could use the instance to send requests and consume the account’s paid resources. It is not a complete Chrome implementation, and some developer-tool functions remain unsupported. This is not a replacement for a phone’s everyday browser. Its immediate appeal is to developers who want to inspect a failure, alter the engine and test whether their particular workflow belongs on it.

Original source published: Cloudflare, November 12, 2026; technical explanation published November 13.