Meteor Forked Chromium and Built an AI-Native Browser From the Wreckage

The Macro: The Browser Has Not Changed in a Decade

Think about what you actually do in a browser on any given day. You open tabs. You read pages. You click through multi-step workflows. You copy information from one tab and paste it into another. You compare prices across sites by switching between tabs. You fill out forms. You wait for pages to load. You search for something, click a result, go back, click another result.

Most of this is tedious. Some of it is actively painful. And almost none of it has changed in any fundamental way since Chrome launched in 2008. Browsers got faster. They got better at rendering web apps. They added tab groups and reading lists and picture-in-picture. But the core interaction model, you navigate, you read, you click, you type, has been static for nearly two decades.

Meanwhile, AI agents can now navigate web pages, interpret content, fill out forms, and execute multi-step tasks autonomously. The capability exists. But the browser itself is not designed for it. When you use an AI agent to do something in a browser today, it is typically running in a headless browser instance or a cloud-based automation environment. The agent works in a separate context from where you work. You cannot see what it is doing. You cannot intervene mid-task. You cannot hand off seamlessly between doing something yourself and letting the agent take over.

The browser competitors in the AI-adjacent space are mostly taking an extension approach. Arc added AI features on top of its browser. Brave has a built-in AI assistant. Various Chrome extensions bolt AI capabilities onto the existing browser. But extensions and add-ons are fundamentally limited by the browser’s architecture. They can inject scripts, modify pages, and add sidebars. They cannot change how the browser itself works at the rendering, navigation, or tab management level.

The harder and more ambitious approach is to build a browser where AI is native to the architecture. Not bolted on. Not injected via extension. Built into the core of how the browser operates. That is what Meteor is attempting.

The Micro: 30 Million Lines of Code, Forked and Rebuilt

Meteor was founded by Pranav Madhukar and Farhan Khan. Pranav built viral campus projects and secured over $100K in university funding before this. Farhan built an IDE for hardware design that compiled 60x faster than industry standards. They came through Y Combinator’s Summer 2025 batch as a two-person team in San Francisco.

The foundational decision that defines Meteor is that they forked Chromium. That is not a trivial thing to do. Chromium’s codebase is over 30 million lines of code. It is the most complex open-source project in the world by many measures. Forking it means inheriting all of that complexity, every rendering quirk, every security patch, every platform-specific behavior. But it also means you get a real browser. Not a toy. Not a wrapper around a WebView. An actual Chromium-based browser that can run every website and every Chrome extension.

They reportedly shipped the initial version within three weeks of their YC batch starting. That is aggressive, and it tells you something about the engineering velocity on the team.

The headline capability is browser agents that score 96.5% accuracy on the WebVoyager benchmark. WebVoyager tests whether an agent can complete real-world web tasks: booking flights, managing emails, navigating complex sites. 96.5% is a high number. For reference, earlier browser automation agents were scoring in the 50 to 70% range on similar benchmarks just a year ago. The gap between 70% and 96% is the gap between a demo and a usable product.

The feature set beyond agents includes a native ad blocker, the ability to chat with any page (ask questions about the content you are looking at without copy-pasting into a separate tool), and app integrations with Gmail, calendar, and social platforms. The ad blocker is a direct shot at Chrome, which has been making ad blocking harder through its Manifest V3 extension changes. Meteor can block ads at the browser level because they control the browser, not through an extension that plays by Chrome’s rules.

The “chat anywhere” feature solves a genuinely annoying workflow. Right now, if I want to ask an AI about something I am reading in my browser, I open a new tab, go to the AI, paste the content, ask my question, and then switch back. It is three or four context switches for what should be a single interaction. Having AI accessible in every tab without leaving the page is the kind of improvement that sounds minor but changes daily behavior.

The Verdict

I have seen a lot of “AI browser” pitches. Most of them are Chrome extensions with better marketing. Meteor is not that. Forking Chromium is a commitment that separates serious attempts from feature experiments. It is also a commitment that creates ongoing maintenance burden, because every Chromium security update needs to be merged downstream.

The 96.5% WebVoyager accuracy is the number that matters most right now. If that holds up in real-world usage and not just on benchmark tasks, Meteor has a browser agent that is reliable enough for daily use. That would be a first. Previous browser agents I have tried (and I have tried most of them) work impressively on demos and then fail on the specific tasks I actually need done.

In 30 days I want to use it daily and see where the agent breaks. Benchmarks are controlled. Real browsing is not. Sites with CAPTCHAs, aggressive bot detection, complex JavaScript frameworks, and authentication flows will stress-test the agent in ways that WebVoyager does not.

In 60 days the question is switching cost. Getting someone to change their browser is one of the hardest things in consumer software. Chrome has decades of muscle memory, saved passwords, bookmarks, and extensions. Meteor needs the agent experience to be so much better that people are willing to move their entire browsing life. Chrome extension compatibility helps, but it is not sufficient.

In 90 days I want to see the retention numbers. Downloads and installs do not matter for browsers. Daily active usage does. If people download Meteor, use the agent for a week, and then drift back to Chrome because the basics are not quite right (rendering bugs, missing features, performance issues), the AI capabilities do not matter.

Building a browser is one of the hardest things in software. Building an AI-native browser is harder. But if any two-person team can make a credible run at it, a duo that forked 30 million lines of Chromium and shipped in three weeks is at least in the conversation.