Paul Klein IV (Browserbase) argues that the browser — specifically a browser-based MCP server — is the practical default interface for AI agents to interact with the legacy web. He demonstrates Browserbase running cloud browsers to automate site interactions (including a live demo that attempts to adopt a dog and shows session replay/DOM), explains how LLMs can generate browser automation code dynamically, and contrasts Browserbase/Stagehand (production framework) with BrowserUse (POC). He closes with a note on remote/asynchronous agents and that the MCP server is available now.
Chapters
Don't look at my files. Okay, I have a confession. I'm a Figma Slides user. There are dozens of us out there. Round of applause. I'm here to talk about how it's not attention, that's all you need. It's the browser, that's all you need. Or the browser MCP server, that's all you need. When I think about browser based, what we do is we're browser infrastructure. We run thousands of browsers in the cloud. For your AI agent to go out and, and connect to the web and do things and solve captchas or use proxies or all the stuff you really need to automate the web with your AI agent. And when I think about AI agents, I think about the legacy Internet. The Internet that's not going to have an MCP server. The DMV will not have an MCP server, at least not under this presidency. Right, so how are we going to have AI talk to the rest of the world? The rest of the Internet? Well, a browser, it's the thing that we all use to interact with the web. So why can't our AI do it too? And I really do believe that the browser is the default MCP server for the rest of the Internet. It's a horizontal MCP server or a primitive. You integrate this one server and you give your agent access to every single website out there, even if there isn't a published MCP available. So of course, live demo gods, we're going to go have our browser do a lot more than take a screenshot. It's going to go out and adopt a dog. So first things first. Can you see that? Okay, great. So we're going to go ahead and create a session in browser base. If I do this right, they should show up next to each other and all my files you can see using the browser based MCP server. It's talking to a browser running in the cloud. First thing was like, oh crap, there's a modal here. Well, I probably got to close that modal and I'm going to go try and find a dog to adopt. Who wants to adopt a dog tonight? Clap your hands really quick, please. That is a ton of dogs we going to be adopting. So we just start with one. You can kind of see it's using tool calls. So with the full context of the HTML as well as the screenshot, models have gotten a lot better at being able to navigate complex pages. Looks like this modal popped up again. They really want that 200k match. So we'll See if the MCD server is able to navigate this and say, oh, actually it did. This is the replay. Sorry it went so fast. So on browser base we actually are able to replay sessions, show you what happened as well as like the DOM of the page and. But this dog they found is pokey. So let's go ahead and pull a pokey really quick. Show me pokey, pokey. Aw. Oh my God. We got to donate tonight, guys. But that's kind of a good. Rejected me. But that's kind of a good example of what the web is like. If you're building deterministic scripts, maybe you've used puppeteer, playwright and selenium before. You know, you've had to hard code these workflows and as soon as a website changes or they roll out the 200k campaign, your automation breaks. With the rise of AI, you can actually have LLMs on the fly generate the browser automation code that actually enables your app to interact with any website dynamically. And that's what browserbase is here to help do. We're an infrastructure company like work os, so our whole thing is that we want to give you the browsers, we want to give the framework, so that's MCP or our framework stagehand to really make automating the web possible. And that's what is really exciting. When we talk to our customers, they're not automating like getting a flight or booking a restaurant. They're automating filing your Delaware franchise tax, or helping you process your immigration in Australia or ordering lumber in Mexico. These are the real problems that people need to solve with AI. And all of you in this room have an opportunity to go do that, especially if you use browser based shout out. Okay, so I have a few more slides here. First of all, I want to talk about the future because I showed you the present. You can go use this today. Browser based MCP we'll talk about in a second. But what's the future? Right now you're running your agent in cursor, right? Maybe this meme should be reversed, honestly, because when I sleep, I want my agent to be running. So I think remote agents or asynchronous agents are on the way. And when you're able to use infrastructure that operates even when your laptop is closed, that's where you're going to see agents really explode. And work OS will be in very good business at that time. So the final thing I'll share is that our MCP server is available today. If you want to go build an agent that can go operate anything on the Internet. You can scan this QR code, install it. It works in cursor, it works in Windsurf, it works everywhere. And that's pretty much it. With 45 seconds to spare. Come on, Michael. I earned. One question, please. All right, round of request. One question. Raise hands. Really quick. How does it matter? Browser use? So browser use is a cloud platform for browsers. Python SDK, Browser base is just the infrastructure. We have Stagehand, our framework. Stagehand has more daily downloads than browser use. And you can think of, like, browser use as more for, like, POCs. Stagehand is for production usage. So if you're going enterprise first or selling to software, consider browser based and sagehand. Cool. Give it up again. Great.
