The road to the agentic browser: A Kitesurf update

Cloudflare has shipped a substantial update to Kitesurf, the browser it introduced in August that runs entirely on Cloudflare Workers and is designed around what AI agents need from the web rather than the feature set of a human-facing browser. The update, described as a step on the road to the agentic browser, adds WebMCP support, expands the set of implemented web standards, tunes the engine for agent-loop latency, completes Browser Run API coverage, and makes Kitesurf runnable from a terminal.
The headline capability is WebMCP. The reasoning behind it is that websites were built for people, so an agent browsing today has to click pixels and hope the right element loads — a slow and fragile approach in a programmatic setting. WebMCP lets developers expose site functionality directly to agents, so an agent can call something like searchFlights() instead of simulating clicks. Cloudflare says it has supported WebMCP since its early days, that Browser Run has supported it when using Chrome beta for some time, and that site owners recently gained a one-switch way to enable it so browser agents can discover and use tools without changing site code. Kitesurf is now added to that list. The company points to its public Kitesurf playground as a test path: open Cloudflare Radar, then WebMCP on the Application tab in the DevTools panel, where Radar exposes tools such as navigate-to and set-location for programmatic interaction. Agents can be pointed at Kitesurf through a chrome-devtools-mcp configuration that sets a WebSocket endpoint scoped to an account and uses the experimental-webmcp category.
In terms of standards coverage, Kitesurf's supported API list has grown to include WebMCP, CSS Layout, the CSS Object Model (CSSOM), CSS Typed OM, and custom elements. The team also added URL-based module resolution, JSON modules, and import map handling — all relevant to sites that load JavaScript in chunks — and is using Cloudflare Workers' new module registry to support imports from URLs. Iframes now load at the right time, stay better isolated, and render text correctly across more languages and encodings.
Progress is tracked through Web Platform Tests (WPT), the shared open-source suite that checks whether browsers implement standards consistently. Kitesurf now passes more than 730,000 WPT subtests, up 500,000 from launch. That is a useful yardstick for developers: WPT counts subtests rather than features, so the number is a rough proxy for breadth rather than a quality score, and passing tests does not guarantee that complex pages behave identically to Chrome.
Efficiency work targets the agentic loop rather than page load time. The stated goal is that agents spend compute cycles on reasoning rather than waiting on the browser. Concretely, JavaScript execution does less work crossing between Boa, Kitesurf's JavaScript engine, and the DOM, making that boundary more compatible with real web frameworks. Common reads such as getAttribute, id, and parentNode can now be answered inside the Wasm DOM rather than making repeated Boa to JavaScript shim to Wasm trips. Kitesurf does less repeated work when running timers and loading scripts, releases memory from objects it no longer needs, and handles objects and classes more consistently across its two JavaScript engines. Font handling is lazier: fonts load when needed, fewer unused fonts are fetched, the browser checks which characters appear on a page before fetching language-specific font files, and synthetic italics render more faithfully. Cloudflare says wall-clock time and CPU usage remain roughly in line with launch benchmarks, and in some cases improved, despite the additional standard support.
Kitesurf now has full Browser Run API coverage. Browser Run is Cloudflare's developer product for programmatically controlling headless browser instances, and Kitesurf is one of the selectable browser flavors. That means Kitesurf can be driven through CDP, Playwright, Puppeteer, or MCP. Quick Actions — simple interfaces for common tasks such as capturing screenshots, extracting HTML, and generating PDFs — were previously available from the REST API with Kitesurf and can now also be called from inside a Worker script through the env.BROWSER.quickAction() binding, with browser set to kitesurf.
Finally, Kitesurf runs in the terminal. Its architecture separates PageScript, the isolate that runs the page session and page code, from PageRenderer, which turns computed page objects into pixels. That split allows rendering to be moved outside Kitesurf — to a client or another Worker — while keeping security-critical work server-side on Cloudflare's network, an approach the company compares to Cloudflare Browser Isolation. To demonstrate it, the team moved PageRenderer to a Playground Worker and adapted it to output to Kitty, a terminal graphics protocol supported by terminals including Kitty, Ghostty, and WezTerm, with a pure ANSI text mode as a fallback where Kitty is unavailable. The terminal accepts scrolling and click events, so keyboard, arrow keys, and mouse work as they would in a dedicated browser window. Installation is via brew install cloudflare/cloudflare/kitesurf, after which kitesurf https://blog.cloudflare.com opens a page. Beyond convenience, this gives developers a way to inspect how an agent using Kitesurf perceives a page.
On roadmap and licensing, Cloudflare says it will publish performance benchmarks and WPT results openly so progress can be tracked in real time, and will keep exploring scenarios where decoupling PageRenderer from PageScript helps agents or where higher frame rates matter. The company reiterates its commitment to open-sourcing Kitesurf, saying it is coming soon but will be done in a way that supports the project long term. Kitesurf continues to run entirely on Workers using only publicly available features and APIs, with no special privileges, which Cloudflare frames as both platform dogfooding and the reason it can be cheap and scale across its network. It is free while in beta, subject to per-account limits.
Why it matters: teams building browser-using agents get a cheaper, standards-oriented target that speaks WebMCP, so sites that expose tools can be driven by function calls rather than brittle pixel clicking. Product teams can already try Kitesurf through the playground, Browser Run with Playwright or Puppeteer, and a terminal CLI without a separate browser application. The main open questions are how the WPT number translates into reliability on demanding real-world pages and what the eventual open-source release and per-account limits look like in production; both are reasonable inferences from a product that is explicitly still in rapid beta.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
We’ve updated Kitesurf, our Workers-based browser for AI agents, with WebMCP support, improved DOM performance, and terminal-based rendering. With over 730,000 Web Platform subtests passing, agents can now navigate complex sites faster.