Cloudflare used its Agents Week to place two bets on the same premise: that a growing share of web traffic is no longer human, and that the infrastructure underneath the web was never designed for it. The company launched a developer preview of WebMCP, which exposes any site to browser AI agents without code changes, and Kitesurf, an agent-first browser that runs entirely in V8 isolates on Cloudflare Workers with no Chromium underneath.
WebMCP: tools instead of scraping
Every design decision in a web page presumes a reader with hands and eyes. That presumption is quietly failing. Where an automated visitor arrives today, the prevailing answer is a crawler that lifts the text back to someone else's server, and the site that produced it sees nothing in return.
WebMCP proposes a different contract. It is a new browser standard, shipping experimentally in Chrome 146, that surfaces in the page as document.modelContext. A site registers a set of tools there, and a visitor's agent calls them directly in the browser. The visit stays on the site, and the human keeps control over what actually gets executed.
Cloudflare's contribution is removing the implementation work. Flip the feature on and the edge inserts a lightweight bridge into outgoing pages, publishing the tools without anyone touching the origin server or shipping a release. Capabilities are bundled into packs that toggle as a unit; the developer preview carries two of them, and both execute client-side. Additional packs become available to a site the moment it enables them.
Kitesurf: a browser sized for agents
Kitesurf attacks a different bottleneck. Rendering engines carry decades of accumulated machinery meant to satisfy human eyes, and that weight is what makes handing every agent its own browser financially absurd once the fleet grows. What an automated client actually wants, by Cloudflare's reckoning, is parseable output, a small token footprint, and a hard boundary around anything it reads — chrome, plugins and smooth animation are dead weight.
So Kitesurf drops Chromium entirely and runs as a stateless browser inside V8 isolates on Workers. Cloudflare reports it uses 3.1 to 3.8 times less CPU and 4.7 to 7.0 times less memory than Chromium on common agent tasks such as capturing screenshots and extracting HTML. It passes more than 215,000 Web Platform Tests, and existing Puppeteer, Playwright and Chrome DevTools Protocol clients can point at it by adding a single browser=kitesurf parameter.
The gaps are disclosed rather than buried. Video playback, WebGL, TLS-fingerprint challenges and sessions that stay logged in across requests are all still missing — a short list that rules out a good deal of real automation work. Browser Run carries it today at no charge for the duration of the beta.
One argument from both ends
Read together, the two launches are the same position stated from opposite directions. WebMCP gives publishers a way to serve agents that keeps traffic and control on their side of the transaction. Kitesurf gives agent builders a cheap, sandboxed way to browse when a site has not opted in.
That position lines up with Cloudflare's crawler policy, which gives AI companies until 15 September to separate crawlers used for search from those used for AI training and agents, or risk being blocked by default across a large number of publisher sites. The company is effectively offering the industry a trade: stop scraping, and it will build the plumbing that makes the alternative work.
Whether publishers adopt WebMCP in volume is the open question. The standard is experimental, browser support is one release deep, and the incentive only holds if agent traffic turns out to be worth capturing rather than blocking.
