Cloudflare open-sources Kitesurf: AI agents get by on 58 MiB of memory instead of Chromium's 399
Cloudflare has open-sourced Kitesurf, a stripped-down browser for AI agents that need to read websites and fill in forms rather than show a picture to a human. To take one screenshot, it needs 58 MiB of memory and 352 ms of CPU time. Chromium (the engine behind Chrome) uses 399 MiB and 1880 ms. The price is speed: 1.51 s vs. 1.21 s. If you run thousands of agents against ordinary pages, this is a way to cut your cloud bill.
AI-processed from Habr AI; edited by Hamidun News
Imagine you have launched a thousand AI agents, and each one has to open a website and pull out text or fill in a form. Usually that means spinning up a full Chrome instance per agent, or more precisely its engine, Chromium, which is built for a live human: it renders images, plays video, and keeps dozens of tabs open. For an agent, that is dead weight, and you pay for it in cloud memory and CPU time.
Cloudflare has open-sourced Kitesurf, an experimental browser engine that does only what an agent needs. To take one screenshot, Kitesurf uses 58 MiB of memory (a mebibyte, nearly the same as a megabyte) and 352 ms of CPU time, while Chromium uses 399 MiB and 1880 ms. The company unveiled the project in August 2026.
What Kitesurf can do
Kitesurf loads web pages, parses HTML, applies CSS, runs JavaScript, and takes screenshots. An agent can navigate between pages, extract content, and fill in forms. Developers can modify the code, run the engine on their own machines, and embed it in their own cloud systems. Cloudflare says this architecture makes such tasks cheaper than they are with Chromium.
Under the hood are Rust modules recompiled to WebAssembly. Blitz handles HTML and page structure, Firefox's Stylo processes CSS, and the Linebender libraries take care of text and 2D graphics. JavaScript runs on a separate engine, Boa, which also runs inside WebAssembly.
The engine can be controlled through Chrome DevTools Protocol (CDP), the standard way to automate browsers. Agents connect through MCP. In late summer, the team added WebMCP support: if a site supports it, the agent directly calls the functions the site itself has described, instead of hunting for buttons and clicking through the page step by step.
The numbers: cheaper, but slower
In the screenshot test, Kitesurf used 352 ms of CPU time and 58 MiB of memory, and total execution time was 1.51 s. For Chromium, the same metrics are 1880 ms, 399 MiB, and 1.21 s. That works out to roughly 5 times less CPU and almost 7 times less memory, but the task takes 0.3 s longer, which makes Kitesurf about 25% slower.
In practice, this means the following. If your agents don't need an answer within a fraction of a second and instead need to process many pages — collecting data, reading documentation, submitting a simple form — the savings in memory and CPU matter more than the extra 0.3 s. A single machine can fit noticeably more parallel sessions, and the cloud bill goes down. If minimal latency matters, Chromium is still faster.
Limitations
The project is experimental, and its compatibility with the modern web is still incomplete. The engine is not suited to video playback, WebGL, or long sessions with authentication. Rendering can also differ from Chromium: a Kitesurf screenshot will not always match what a user sees in a regular browser. If you check layouts down to the pixel, this tool is not for you.
What to do with it
The code is available on GitHub, and the documentation is published on the project's website. A sensible plan for a team running agents: take a set of your most frequent tasks, run them through both Kitesurf and Chromium, and compare what share of pages produce matching results. Simple scenarios such as reading pages, extracting text, and filling in basic forms can move to the lightweight engine. Sites with video, WebGL, and long account logins are better left on a full browser.
There is also a broader takeaway. If WebMCP catches on, part of the agents' work will shift from "look at the page and click" to direct calls to a site's functions. A heavy browser would then be needed less often, and lightweight engines like Kitesurf could become a convenient foundation for agents running at scale.
- Nov 18, 2026OpenAI agent fills out insurance paperwork: Luna scores 91.8% for 8.5 cents
- Nov 18, 2026Zeroing Out Less Than 0.05% of Weights Taught a Large LLM New Skills as Well as Fine-Tuning
- Nov 18, 2026MoEngage Plugs AI Agents Into Its Existing APIs: Glance's Push CTR Up 38%
- Nov 18, 2026Hark makes its Handoff² browser agent twice as fast: task success rises from 73% to 82%
Key Questions Answered
What Kitesurf can do?
Kitesurf loads web pages, parses HTML, applies CSS, runs JavaScript, and takes screenshots. An agent can navigate between pages, extract content, and fill in forms. Developers can modify the code, run the engine on their own machines, and embed it in their own cloud systems. Cloudflare says this architecture makes such tasks cheaper than they are with Chromium. Under the hood are Rust modules recompiled to WebAssembly.
What are the key details on The numbers: cheaper, but slower?
In the screenshot test, Kitesurf used 352 ms of CPU time and 58 MiB of memory, and total execution time was 1.51 s. For Chromium, the same metrics are 1880 ms, 399 MiB, and 1.21 s. That works out to roughly 5 times less CPU and almost 7 times less memory, but the task takes 0.3 s longer, which makes Kitesurf about 25% slower. In practice, this means the following.
What are the key details on Limitations?
The project is experimental, and its compatibility with the modern web is still incomplete. The engine is not suited to video playback, WebGL, or long sessions with authentication. Rendering can also differ from Chromium: a Kitesurf screenshot will not always match what a user sees in a regular browser. If you check layouts down to the pixel, this tool is not for you.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.