Cloudflare has expanded the capabilities of Kitesurf, its browser built on Cloudflare Workers and designed to run AI-agent tasks, by adding WebMCP support, improvements to the browser engine, and full integration with the Browser Run APIs. It is also now possible to run a version of Kitesurf in the terminal, reflecting its design, which differs from traditional browsers primarily aimed at human users.
WebMCP shifts interaction from clicking to tool calls
WebMCP allows websites to expose their functions directly to agents, enabling an agent to call tools such as searchFlights() instead of simulating clicks on elements and waiting for pages to load. Cloudflare says Kitesurf currently supports WebMCP, and this can be tested through the public Kitesurf environment and Cloudflare Radar, which exposes tools such as navigate-to and set-location for programmatic interaction with the page.
Using these capabilities with an AI agent requires setting up a connection through Chrome DevTools MCP and the Browser Run WebSocket endpoint, while passing account credentials. Cloudflare provides separate documentation for using WebMCP with Kitesurf and Browser Run.
Broader coverage of web standards
Cloudflare has added support for WebMCP, CSS layouts, CSSOM, CSS Typed OM, and custom elements. Kitesurf also now handles module resolution through URLs, JSON modules, and import maps, making use of the Cloudflare Workers module registry to support imports from web addresses.
The improvements include iframes, which now load at the appropriate time with better isolation and more consistent text rendering across languages and character encodings. According to Cloudflare, Kitesurf now passes more than 730,000 Web Platform Tests subtests, an increase of 500,000 tests compared with the launch stage.
Improvements targeting agent-loop latency
Cloudflare is focusing on reducing the time an agent spends repeatedly interacting with the browser, rather than only on page-load speed. The changes include reducing transitions between the JavaScript and DOM engines, allowing common operations such as getAttribute, id, and parentNode to run inside the Wasm DOM instead of repeatedly passing through intermediary layers.
Kitesurf has also reduced repeated work during timer execution and script loading, and improved memory management and object handling as objects move between the JavaScript engines. Font loading is now more selective, with fonts fetched only when needed after checking the characters used on the page. The company says CPU consumption and wall-clock time remained close to launch measurements and even improved in some cases despite the addition of support for new standards.
Operation through Browser Run and the terminal
Kitesurf now has full coverage of the Browser Run APIs and can be used through CDP, Playwright, Puppeteer, and MCP. Actions such as taking screenshots, extracting HTML, and creating PDF files are also available from within a Worker through the env.BROWSER.quickAction() interface.
In the terminal, Cloudflare separated the PageRenderer component responsible for producing the image from PageScript, which manages the page session and executes its code. The new version uses the Kitty graphics protocol, with an ANSI-based text mode for environments that do not support Kitty. It can be installed through Homebrew and then launched using the kitesurf command, with support for scrolling, clicking, and the keyboard.
What changes in practice?
The update combines greater programmability for websites through WebMCP, broader compatibility with web applications, and operating modes suited to developers and agents. However, Kitesurf remains in beta and is available for free within per-account limits. Cloudflare’s announcement that it intends to open-source the project in the future still does not include a specific date, making the scope of current availability and the continuity of the interfaces points that require monitoring.