Cloudflare adds WebMCP, terminal rendering and 730,000 WPT passes to its agent browser Kitesurf

Cloudflare has published a progress update on Kitesurf, the experimental browser it designed for AI agents rather than people, announcing WebMCP support, much broader web standards coverage, full Browser Run API integration and a terminal client that can render pages as graphics or plain text.

According to the Cloudflare blog, Kitesurf was introduced in August and runs entirely on Cloudflare Workers. The company says it has since been tested on increasingly realistic tasks and refined using internal and external customer feedback.

WebMCP support

Cloudflare points out that websites were not built for agents: browsing today is a process of clicking pixels and hoping the right element loads, which is slow and fragile in a programmatic world. WebMCP addresses this by letting developers expose site functionality directly, so an agent can call a function like searchFlights() instead of simulating clicks.

Cloudflare says it has supported WebMCP since its early days and recently announced a one-switch option that lets site owners expose tools to browser agents without changing site code. Kitesurf itself now supports the protocol. In the public Kitesurf playground, Cloudflare Radar exposes WebMCP tools such as navigate-to and set-location, discoverable in the WebMCP section of the DevTools Application tab. Agents can target Kitesurf through chrome-devtools-mcp with an experimental WebMCP flag, which Cloudflare says lets models complete tasks more reliably.

Broader standards coverage

Since launch, Cloudflare has added APIs including CSS Layout, CSS Object Model, CSS Typed OM and custom elements, plus URL-based module resolution, JSON modules and import map handling, which matters for sites that load JavaScript in chunks. Imports from URLs are handled through a new Cloudflare Workers module registry. Iframe behavior also improved: frames load at the right time, stay better isolated and display text correctly across more languages and encodings.

On Web Platform Tests, a shared open-source suite that checks consistent standards implementation, Kitesurf now passes more than 730,000 subtests — roughly 500,000 more than at launch, according to Cloudflare.

Efficiency tuned for the agentic loop

For an AI agent, Cloudflare argues, efficiency is about the latency of the agentic loop rather than how fast a page appears. The team says it aggressively optimized the engine internals so every DOM traversal, timer and font fetch is as lightweight as possible, leaving the agent's compute for reasoning instead of waiting on the browser.

Among the changes: common reads such as getAttribute, id and parentNode can now be answered inside the Wasm DOM rather than making repeated trips between the Boa JavaScript engine and the DOM, making the boundary more compatible with real web frameworks. Timers and script loading do less repeated work, memory is released from objects no longer needed, and objects and classes are handled more consistently when code moves between the two JavaScript engines. Fonts now load on demand, with the engine checking which characters appear on a page before fetching language-specific files and rendering synthetic italics more faithfully. Despite the added standards support, wall-clock time and CPU usage remain roughly in line with launch benchmarks and have improved in some cases, Cloudflare reports.

Full Browser Run API coverage

Browser Run is Cloudflare's product for programmatically controlling headless browser instances, selectable from a list of browser flavors that includes Kitesurf. According to the post, Kitesurf now has full Browser Run API coverage and can be used with CDP, Playwright, Puppeteer or MCP.

Quick Actions, popular interfaces for tasks such as capturing screenshots, extracting HTML and generating PDFs, were previously available only through the REST API. They can now also be invoked from inside a Worker script via an env.BROWSER.quickAction() binding.

A browser in the terminal

Kitesurf separates PageScript, the isolate that runs page code, from PageRenderer, which generates pixels from computed page objects. Cloudflare says this decoupling allows rendering logic to move outside the browser — to a client or another Worker — while security-critical parts stay server-side, a model comparable to its Cloudflare Browser Isolation SASE product, which streams rendering data from the edge.

To prove it, the team moved PageRenderer into its Playground Worker and patched it to output through Kitty, a terminal graphics protocol supported by terminals like Kitty, Ghostty and WezTerm, with a pure ANSI text mode as a fallback. Pages can now be rendered without leaving the terminal via brew install cloudflare/cloudflare/kitesurf, which Cloudflare also frames as a way to see how an agent using Kitesurf perceives a page.

Why it matters

Most agent browsing today drives human-oriented browsers and inherits their weight and fragility. Kitesurf tests the opposite bet: a browser whose API surface, efficiency profile and rendering model are shaped around agents. The WebMCP support is the most consequential piece — if sites expose callable functions and agent browsers can discover them, the era of simulating clicks could start to fade. The 730,000-plus WPT passes show Cloudflare closing the compatibility gap with full-featured browsers like Chrome, and the terminal renderer hints at rendering as a service for machines rather than screens. It is still early, but the update sketches a plausible shape for the emerging agentic web.