Cloudflare Kitesurf: The Agent-First Browser on Workers

Kitesurf is a stateless browser built exclusively for AI agents, announced by Cloudflare on August 6, 2026. It uses no Chromium at all — it's an entirely new browser engine that runs purely inside V8 isolates on Cloudflare Workers, built around the concept of "a browser for the Agentic Cloud." During the beta period it's available for free via Browser Run (Cloudflare's browser execution service), with per-account usage limits. Full details are covered in Cloudflare's official announcement.

Traditional browsers like Chromium are designed for humans. Tab management, themes, extensions, 60fps smooth scrolling, and pixel-perfect rendering are essential for human users, but they're pure overhead for AI agents. In its announcement, Cloudflare argues that what agents actually need is machine-readable content, low token overhead, scalability, and isolation from threats such as prompt injection. Kitesurf was designed specifically around these four requirements. It runs on the same kind of Workers-based edge execution environment described in our guide to building edge APIs with Hono and Cloudflare Workers, which is well suited to spinning up lightweight isolates per request.

Kitesurf is built in Rust and WebAssembly, and consists of four components:

- Engine: manages the Chrome DevTools Protocol (CDP) connection and session state
- PageScript: runs JavaScript/WebAssembly on Dynamic Workers. HTML parsing uses the modular rendering engine Blitz, CSS parsing uses Firefox's Stylo, and eval() support is handled by the Rust-based JS engine Boa
- PageRenderer: produces rasterized output such as screenshots using blitz-paint
- SandboxOutbound: isolates network access, enforcing CORS and managing cookies on a per-page basis

Each page navigation is capped at a maximum of 20 seconds of CPU time and 60 seconds of wall-clock time; processing is automatically halted if either limit is exceeded.

Cloudflare has published benchmark results comparing Kitesurf against Chromium under warm-pool conditions.

The results reveal a clear trade-off. CPU and memory consumption are 3-7x lower than Chromium — numbers that translate directly into usage-based billing costs. On the other hand, wall-clock time for completing a screenshot or HTML extraction is 1.7-1.8x slower, since Kitesurf doesn't benefit from Chromium's JIT optimizations. In short, Kitesurf trades "fast but expensive" for "somewhat slower but cheap." It has also passed over 215,000 Web Platform Tests, giving it solid coverage across CSS, DOM, HTML, SVG, and XHR.

There are three main ways to use Kitesurf.

The first is through API/MCP clients such as Puppeteer or Playwright: simply add the browser=kitesurf parameter to an existing Browser Run connection. Via MCP, it's also accessible from clients like Claude, Cursor, and Gemini CLI.

The second is Quick Actions, for serverless screenshot or PDF generation.

The third is the official playground, which lets you try it directly in the browser with Chrome DevTools inspection built in. Supported output formats include inspection data, PNG screenshots, PDF, and HTML.

Kitesurf has clearly documented limitations at this stage. It doesn't yet support video playback, WebGL, TLS fingerprinting for bot challenges, or long-running authenticated sessions that require persistent state. For workloads that need these, Browser Run's default Chromium remains available as a parallel option. Here's a rough guide to choosing between them:

The roadmap includes expanding CDP coverage, improving rendering fidelity, raising WPT pass rates, further performance optimization, and open-sourcing the project. In terms of state management for edge execution environments, it's likely we'll see Kitesurf increasingly paired with persistence mechanisms like those described in our explainer on Cloudflare Durable Objects.

For teams running web scraping, E2E monitoring, or AI agent infrastructure, the cost structure of browser execution maps directly onto CPU and memory usage-based billing. For workloads that run agent processes at high volume and in parallel, the potential to cut CPU and memory usage by 3-7x compared to Chromium can meaningfully affect infrastructure costs. At the same time, the 1.7-1.8x slower wall-clock time means it may not suit workloads that demand real-time responsiveness. Deciding when to use Kitesurf versus traditional Chromium, on a case-by-case basis, is likely to become an increasingly important design decision in contract development and systems operations going forward.

Kitesurf abandons the human-oriented assumptions baked into Chromium and is designed purely for how AI agents actually use a browser. It offers a clear win on the metrics that drive billing — CPU and memory efficiency — while officially disclosed benchmarks show it trades that off against slower wall-clock time, a nuance worth understanding before adopting it. Since it's free to try during the beta, engineers working on scraping, E2E monitoring, or AI agent infrastructure may find it worthwhile to test its behavior directly on the official playground.