🌐 Two Years Later: An Internet Browser for AI

A follow up on the memo that started my company.

George Callaway Matelich was born on October 22nd, 2025. I can’t wait to read his memos one of these days. Congratulations George, Anne and their families!

Two years ago, I wrote An Internet Browser for AI. I didn’t know it at the time, but that memo would change my life. It became the foundation for Browserbase and set me on a path that still feels both inevitable and unpredictable.

When I wrote it, I remember thinking how neatly my background came together for this moment. I started in infrastructure as an engineer at Twilio, learning how large-scale systems actually run. At Stream Club, I was a founder working directly with headless browsers, building products on top of them and understanding their quirks. Then at Mux, I led teams, worked on developer tools, and spent time on developer marketing, learning how to connect infrastructure products with their users. Those three experiences (engineering, founding, and communicating) fit together in a way that made this company feel like the obvious next step in hindsight.

When I finished the memo, I sent it to a few people whose opinions I trusted most. They’d seen earlier drafts of my thinking before, but this time the reaction was different. Everyone agreed that the idea was strong and original. Some people even wanted to invest. That was when I realized the memo wasn’t just a piece of writing, it could actually become a company.

Predictions and Risks

Looking back, most of the predictions in that first memo proved right.

I wrote that AI would need to use the internet the same way people do, and that browsers would become one of the core tools for it. That’s true today. Web agents exist in many different forms and already power a wide range of applications.

I said scraping and browser automation would remain essential even as new access methods appeared. That’s still the case. Even with the rise of MCP, there is no escaping the need to interact directly with websites.

I argued that existing tools were suboptimal, and I still believe that. Playwright has improved a lot and deserves credit, but Stagehand continues to be faster and more token efficient.

I also believed that open source, community, brand, and a bottoms-up strategy were the right way to build in this market. Two years later, that’s worked. Stagehand’s adoption has grown faster than I expected, and our brand has become one of Browserbase’s biggest strengths.

The broader category has expanded dramatically. Playwright’s installs have multiplied since I wrote the memo, and browser automation has become a standard developer primitive.

Some of the risks I listed turned out to be smaller than expected. Models improved at an unbelievable rate. Inference costs dropped steadily. Legal and ethical fears that could have slowed the space haven’t materialized. If anything, the tone of the conversation around AI has become more constructive and forward-looking.

The core thesis, that browsers would remain essential, has only strengthened. The internet is not being rebuilt for AI anytime soon. The long tail of websites still needs to be read, clicked, and filled out, and that means there will always be a need for a browser layer built for automation.

MCP and the Long Tail of the Web

When I wrote the original memo, I didn’t predict that MCP would exist, let alone that it would take off. I believed AI would need to use websites and tools, and that websites would make up part of that toolset. What I didn’t expect was that there would be a standardized protocol for connecting AI systems to tools at all. That part surprised me.

MCP has turned out to be a clever idea. It gives developers and agents a shared language for connecting to external capabilities. That’s a big step forward for interoperability. But it doesn’t replace the web. If APIs didn’t eliminate the need for browsers, MCP won’t either. It’s mainly a wrapper around APIs and serves a complementary purpose.

Most of the internet will never have an MCP server. The same organizations that haven’t built APIs are unlikely to build these either. That long tail of messy and outdated websites is where AI still needs to operate. That’s the part of the internet that Browserbase is built for.

Customers and Market Reality

Our customer base has evolved alongside the category’s maturity.

In the early days, most of our users were bleeding-edge startups. They were small, fast, and experimenting with AI in ways that others hadn’t yet imagined. They helped us learn faster than any large enterprise could.

Those early builders pushed the limits of the models and created the patterns that defined how people use browser automation today. Features like utilizing accessibility trees and the syntax that made Stagehand approachable came directly from watching how those customers worked.

That created a product-adoption flywheel that still drives us. Bleeding-edge customers had new problems, which pushed our roadmap forward. Solving those problems made our product better than anything else on the market. That, in turn, attracted more customers with new and harder problems. The cycle kept repeating, and the product kept improving.

Things began to change after OpenAI launched Operator in early 2025. That was the moment people outside our corner of the world realized that AI could actually use the browser. Suddenly, the concept clicked. From that point, we started moving upmarket—mid-market and enterprise customers, including some very recognizable names, began adopting Browserbase.

I still love working with the early builders, but now there’s a new kind of energy coming from the larger companies figuring out how to deploy this technology at scale.

The Next 24 Months

The next big shift will come when browser agents stop being treated as toys that only do simple things like automating LinkedIn. When they start performing complex tasks reliably, people will realize how capable they are.

I think browser agents will eventually feel as powerful as humanoid robots. Today, people look at them the way they look at Roombas: limited, clunky, useful in narrow ways. Over the next couple of years, that perception will change. These systems will start to feel extremely capable.

We’ll know the inflection has arrived when people stop talking about “browser agents” as something new and start treating them as a normal part of software development. It will feel obvious, the same way AI coding assistants feel now.

It’s hard to predict exactly when that happens. Maybe 2026, maybe sooner. But I’m certain it’s on the way.

Reflections

I wrote the original memo sitting on my couch, watching football, blissfully unaware of how much life was about to change.

I’ve learned a lot in these past two years. Every Monday, I walk into the office wondering what fire I’ll have to put out that week. The thing you learn is that you always find a way through it.

If I could tell my 2023 self one thing before publishing that memo, it would be to trust my instinct. The idea was right, and the timing was even better than I realized.

Two years later, the browser isn’t just a tool for humans. It’s becoming a universal interface for intelligence itself. It connects the digital world in ways no other layer can. And we’re still just getting started.