Cloudflare has launched Kitesurf, a cloud-hosted browser designed specifically for artificial-intelligence agents rather than human users. The new tool is intended to let software agents navigate websites, fill out forms, extract information, generate screenshots and complete browser-based tasks without developers having to build and operate their own browser infrastructure.
Kitesurf is not a consumer alternative to Google Chrome, Apple Safari, or Microsoft Edge. It has no tabs, themes, extensions, or polished visual interface for people. Instead, it is an agent-first, stateless browser built to operate at scale inside Cloudflare Workers, the company’s serverless computing platform.
Cloudflare says the browser uses substantially less CPU and memory than Chromium for common automation tasks. The trade-off is speed and fidelity: Kitesurf is slower in wall-clock time on Cloudflare’s benchmarks and does not aim to reproduce pages with the pixel-perfect visual accuracy of a full-featured desktop browser.
The product is available free in beta through Cloudflare’s Browser Run service, though usage remains subject to per-account limits.
A browser for machines, not people
Traditional browsers are built around human expectations. They maintain tabs, render web pages smoothly at high frame rates, support extensions, save passwords, preserve browsing state, and display complicated page layouts with close visual precision.
AI agents do not need many of those features.
An AI agent generally needs to open a page, understand its contents, find an element, fill out a form, capture a screenshot, extract structured data, navigate to another page, or return a result to a model. Its priorities are typically different from those of a person browsing the web:
- Lower CPU and memory consumption
- Fast startup for short-lived tasks
- Clear HTML and DOM access
- Strong isolation between sessions
- Low operating costs at high volume
- Compatibility with automation frameworks
- Manageable context windows and token costs
Kitesurf is built around those needs. Cloudflare’s documentation says the system focuses on “token count, context windows, scalability, performance, and cost,” while intentionally trading away features primarily useful to human browsers, including themes, tabs, extensions, and pixel-perfect rendering.
The design reflects a broader shift in the AI industry. As companies develop agents that can browse the web on behalf of users, browser infrastructure is becoming a core part of the AI stack. A chatbot can answer a question from its training data, but an agent that needs current information or must complete a real-world task may need to interact with the web directly.
That activity can be expensive. Running thousands of Chromium instances to support AI agents consumes significant memory and computing resources. Cloudflare’s pitch is that a stripped-down, cloud-native browser can make those workloads cheaper and easier to scale.
Built on Workers, not Chromium
Kitesurf runs entirely on Cloudflare Workers and does not rely on Chromium underneath.
Workers is Cloudflare’s serverless computing environment, which allows developers to run code across the company’s global network without maintaining traditional servers. Kitesurf uses that environment to create short-lived browser sessions that can start when an agent needs them and disappear after the task is complete.
At the technical level, Kitesurf runs in V8 isolates, lightweight execution environments associated with the JavaScript engine that powers Chrome and other systems. Cloudflare says the browser uses a modular rendering engine from Blitz, Stylo, the CSS parser used by Firefox and Boa, a Rust-based ECMAScript engine. Those components are compiled to WebAssembly and run inside Workers.
The result is a browser architecture that differs sharply from the traditional model. Instead of launching a large browser process with the broad range of features needed for human interaction, Kitesurf creates an ephemeral environment optimized for a specific agentic job.
Cloudflare said it built Kitesurf in about 12 weeks.
How developers use it
Kitesurf is integrated into Browser Run, Cloudflare’s service for programmatically controlling headless browser sessions across its network. Developers can use it through the Chrome DevTools Protocol, or CDP, which means common automation tools can communicate with it.
That compatibility is significant because many developers already use tools such as:
- Puppeteer
- Playwright
- chrome-remote-interface
- Model Context Protocol, or MCP, clients
Cloudflare says existing clients do not need to be rewritten. Developers can switch Browser Run requests to Kitesurf by adding a browser=kitesurf parameter to the relevant endpoint.
For example, an AI research agent could use Kitesurf to visit a set of websites, capture the HTML, identify relevant passages, and return those passages to a language model for analysis. A customer-service agent could navigate a web portal, look up information and populate a form. A monitoring tool could take scheduled screenshots of a dashboard or generate PDF records of a page.
Kitesurf is particularly designed for “one-shot” jobs, which Cloudflare calls Quick Actions. These include extracting page content, producing a screenshot or rendering a PDF for a compatible website.
The efficiency claim
Cloudflare’s central argument for Kitesurf is efficiency.
In its own tests, the company compared Kitesurf with a warm pool of Chromium browsers across a corpus of 14 URLs. It found that Kitesurf consumed significantly less CPU and memory on screenshots and HTML extraction.
| Task | Kitesurf | Chromium | Kitesurf result |
| CPU for screenshot | 380 ms | 1,173 ms | 3.1x less CPU |
| CPU for HTML extraction | 229 ms | 877 ms | 3.8x less CPU |
| Memory for screenshot | 57.8 MiB | 271.0 MiB | 4.7x less memory |
| Memory for HTML extraction | 39.4 MiB | 273.7 MiB | 7x less memory |
The savings are substantial in theory. If a company runs thousands of AI-agent sessions every hour, a browser using one-seventh the memory for HTML extraction could reduce infrastructure costs and allow far more concurrent tasks on the same underlying compute resources.
But Cloudflare’s benchmark also shows a meaningful downside: Kitesurf took longer to finish the tasks.
For screenshots, Kitesurf’s wall time was 1,148 milliseconds, compared with 637 milliseconds for Chromium, making it about 1.8 times slower. For HTML extraction, it took 820 milliseconds compared with Chromium’s 472 milliseconds, or about 1.7 times slower.
Cloudflare’s claim is therefore not that Kitesurf is universally faster. It is that it is more computationally efficient, using fewer resources even when it takes somewhat longer to complete a task.
That distinction matters for developers. A real-time customer-facing agent may value low latency and choose Chromium for some jobs. A high-volume background system that needs to process many pages cheaply may find Kitesurf more attractive.
Isolation and security
Cloudflare is also marketing Kitesurf as a security-focused browser environment for AI agents.
Each page is designed to be isolated, reducing the chance that data from one agent task leaks into another. The company describes Kitesurf as stateless and ephemeral: a session exists only for the duration of a task and does not retain long-term browsing state by default.
That architecture is important because AI agents face unusual web-security risks.
An agent browsing the internet may encounter malicious instructions hidden in webpages, deceptive forms, compromised scripts, or prompt-injection attempts intended to manipulate its behavior. If an agent opens untrusted pages while holding access to authenticated accounts, private documents or sensitive tools, a security failure could expose data or trigger unwanted actions.
Kitesurf does not solve every version of that problem. But isolating individual page loads and avoiding persistent state can limit how far a malicious page can affect other browser contexts.
Cloudflare says the browser treats each page load as untrusted input and isolates workloads so information does not spill between sessions.
What Kitesurf can render
The beta browser already supports a large amount of ordinary web content. Cloudflare says Kitesurf passes more than 215,000 Web Platform Tests and reports strong coverage in several core browser areas:
- DOM: 97% subtest coverage
- HTML: 96%
- Selection: 99%
- SVG: 97%
- Encoding: 99%
- CORS: 95%
- XHR: 95%
- URL: 83%
Cloudflare says Kitesurf can correctly render sites and applications including TodoMVC implementations, Wikipedia, Hacker News, the Cloudflare blog and much of Cloudflare’s dashboard.
That makes the browser viable for many agent tasks involving structured data, ordinary pages, simple applications, and standard web forms.
What it cannot do yet
Kitesurf remains a beta product, and Cloudflare is explicit about its limitations.
It is not yet the right option for tasks that require:
- Video playback
- WebGL graphics rendering
- Long-running authenticated sessions with persistent state
- Bot-challenge handshakes requiring real browser TLS fingerprints
- Pixel-perfect visual rendering
These restrictions mean it cannot simply replace Chrome for every agent use case.
A travel-booking agent that must stay signed in for a lengthy session, handle complex anti-bot systems and interact with heavily dynamic webpages may still need Chromium or another full browser. An AI agent that needs to play video, inspect a 3D model, or use GPU-heavy web applications will also need another option.
Cloudflare’s own guidance recommends Kitesurf for compatible websites and stateless, bursty workloads, while treating Chromium as a fallback for complex pages.
Why Kitesurf matters
Kitesurf is a bet that the future of browsers may include two very different product categories: browsers built for people and browsers built for AI.
For decades, browser competition focused on user-facing features, speed, tabs, privacy tools, extensions, design, and compatibility. But AI agents need a different kind of browser: one that can launch quickly, run cheaply, expose structured page information, and handle many parallel tasks securely.
Cloudflare is well-positioned to make that bet. Its global network, Workers platform and existing role in web infrastructure give it an advantage in deploying browser sessions close to users and websites without requiring customers to operate their own browser fleets.
The company plans eventually to open-source Kitesurf and allow customers to deploy their own instances in their own accounts, according to The Next Web.
For now, developers can try the browser free during its beta period through Browser Run.
Kitesurf’s success will depend on whether its efficiency benefits outweigh its limitations. If AI agents become a standard way for software to interact with the web, developers will need reliable browser infrastructure that is cheaper and safer than launching full Chromium instances for every task. Cloudflare is betting that a browser designed from the beginning for machines, not people, will become part of that infrastructure.
