Cloudflare has released Kitesurf under the Apache 2.0 license, putting the source code for its AI-agent browser on GitHub. Developers can inspect it, modify it, run it on a local machine, and control it through the Chrome DevTools Protocol instead of launching a full Chromium browser.[1][2]
That makes Kitesurf immediately interesting for a home AI agent that needs to read pages or take screenshots without dragging an entire desktop browser behind it. It does not yet make Kitesurf a drop-in Chromium replacement. Cloudflare calls the engine experimental, its compatibility remains incomplete, and the company’s own benchmark shows a trade: much lower memory use, but slower rendering in the median test.[2][3]
A browser engine without the browser container
Kitesurf does the basic jobs an agent needs from a browser engine: it fetches HTML, applies CSS, runs page JavaScript, lays out the result, and renders screenshots. It exposes a CDP endpoint, so existing clients such as Puppeteer can connect over a WebSocket and drive it with familiar automation commands.[1][2]
Underneath, it is not a stripped Chromium build. Kitesurf assembles open-source Rust components compiled to WebAssembly: Blitz handles the DOM, layout, and painting; Stylo handles CSS; Boa can execute JavaScript when Cloudflare’s Dynamic Workers cannot; and Linebender libraries plus Krilla cover graphics, text, and PDF output.[1]
That architectural split is the appeal. A local agent usually does not need a tab bar, extensions, password syncing, or all the machinery of a consumer browser. It needs a dependable way to open a page, inspect it, interact with it, and return useful state to a model. Kitesurf is an attempt to build around that narrower job.
It is also real software rather than a future promise. The public repository contains the engine, renderer, CLI, tests, and build files, and Help Net Security separately reported that developers can download and run it now.[2][5]
The memory result is promising; the speed result is not
Cloudflare’s benchmark dashboard compares Kitesurf with its managed Browser Run service and a cold Chrome start across 38 pages, using the median of five runs for each page. For screenshot work, the aggregate median reports 49 MiB of memory for Kitesurf and 330 MiB for cold Chrome.[3]
That is a large reduction in this particular company-run workload. It is the clearest reason for a home server operator to care, especially if several agent sessions must coexist on a machine whose memory is also holding a local language model.
The rest of the chart is less flattering. Kitesurf took a median 2.39 seconds of wall time and 843 milliseconds of CPU time, against 999 milliseconds and 772 milliseconds for cold Chrome.[3] In other words, the published evidence supports “leaner on memory,” not “faster than Chrome.” Cloudflare’s page-by-page results also vary sharply, and at least one page records a Kitesurf error.[3]
HighVRAM has not reproduced these tests. They are first-party results from Cloudflare, not an independent review, and they do not establish how Kitesurf behaves on a reader’s particular mix of dashboards, shops, login pages, or bot-protected sites.
Long local runs need care too. The repository warns that stopped page isolates can accumulate in memory. In one project-reported test, five parallel connections completed 940 page loads and stayed below 4 GB after the development command exposed V8 garbage collection; without that flag, the runtime ran out of memory after roughly 300 pages.[2] That is useful engineering documentation, but it is not a general RAM requirement. Cloudflare publishes no minimum host CPU or memory specification.
Local is free; hosted self-deployment is not
The local setup is straightforward by browser-engine standards, but it is not a one-click application. The current requirements are Bazelisk, Node.js 22 or newer, pnpm 9.15.4, and GNU tar on macOS. Bazel downloads the pinned Rust toolchain and WebAssembly build dependencies; a separate Rust setup is needed only if a developer wants to edit the vendored crate patches.[2]
After cloning the repository and installing dependencies, pnpm run dev builds the WebAssembly modules and starts Wrangler on 127.0.0.1:8787. A Puppeteer client or agent harness can then connect to ws://127.0.0.1:8787/devtools/browser.[1][2]
That local development path does not require a paid Cloudflare account. Running a self-managed hosted instance on Cloudflare does. Kitesurf uses Dynamic Workers to isolate pages, and Dynamic Workers are part of the Workers Paid plan.[1][2] So this is open-source software that can run on a home machine, but its documented production deployment path remains tied to Cloudflare’s infrastructure and pricing.
Compatibility is the deciding limitation
Cloudflare’s Web Platform Test dashboard looks impressive at first glance: Kitesurf passes 98.3% of the subtests it runs. The more revealing figure is that 77.3% of enrolled test files pass, while 56,854 files are skipped—56,823 because they are not enrolled in the dashboard’s tested set.[4]
Some relevant areas are visibly unfinished. The November 16 results show 74.0% of enrolled Fetch subtests passing, 68.0% for Shadow DOM, 58.3% for WebAssembly, and 80.3% for WebMCP.[4] Kitesurf’s own README is blunt that it provides a CDP endpoint rather than a full Chrome implementation and that some Chrome DevTools features do not work yet.[2]
For a controlled workflow—open known documentation, extract rendered text, capture a screenshot, or operate a simple internal tool—that may be acceptable today. For a general-purpose agent expected to wander across arbitrary websites, Chromium remains the safer default. The missing behavior will not always fail neatly; it may produce a page that loads but behaves differently enough to confuse the agent.
The practical next step is a trial, not a migration. Point one non-critical CDP workflow at a local Kitesurf instance, measure its steady-state memory across repeated navigations, and keep Chromium as the fallback. Kitesurf has already made agent browsing more inspectable and potentially much lighter. Reliability on the messy web will decide whether it becomes the browser engine a home agent actually trusts.