BrowserBase Founder on Why Headless Browsers are the Future of AI - Paul Klein IV

The majority of software is online and we've made. We actually like have timed it really well where we have made a lot of the shift from installed software on your desktop to software running on the web somewhere. And I do think the majority of like productive work that AI will automate will happen in a web browser.

All right, Paul, thank you so much for joining the collective nation podcast. Would love just to start off with you just explaining like, what is a headless browser? What was a core problem that you're trying to solve with browser base?

Yeah, Thomas, I'm super stoked to be here. I love the pod. What is a headless browser? And maybe even a higher level. What is browser base? When you think about how computers navigate the Internet, they need to use it the way people do. They have to go to websites and load them. They have to click on buttons the same way we do. Not everything is available in more of like a API native or computer native way. So a headless browser is the same browser that you and I run on our laptops, but running in a server in the cloud somewhere. And just as Chrome is probably one of the beefier applications running on my MacBook Air, it's really hard to run in production at scale. Running tens of thousands of these in a cost effective and reliable way is quite challenging. And that's where my company Browser Base comes in. We handle the hard parts of running headless browsers or browsers running in the cloud at scale for hundreds of different AI companies and also non AI companies that are doing more traditional process automation using a web browser.

Okay, so maybe we can dive into your. So you've written some blogs on sort of exposing the internal strategy of the problem that you're solving. And I think it was so well written that I shared it with friends, family. I loved it because it was like very sort of facts based. You move sort of point to point so you'd make a point and then you support it with some hacker news post explaining exactly what you mean. Or you explain it with some blog posts. So maybe you could talk about existing browser automation tools. They're bloated, they're brittle. What's the single most overlooked technical breakthrough that sets browser Base apart? And then why do you think it's just deeply resonated with developers? Particularly, I know you talked about in the past, just basically writing a blog post, recording a video and getting a lot of traction from just the thought piece itself. Why do you think that resonated with developers?

Man, how long do we have? I can go for Hours on this, right? It's hard to summarize in maybe like a two minute sound bite, but at a high level it's really been a death by a thousand cuts. For developers who need to write code that interacts with the Internet. First and foremost, there's like three top frameworks. Puppeteer, Playwright and Selenium. All of these frameworks were built for testing websites. They take code and then they actually execute commands on a browser. They're browser automation frameworks. And yeah, these frameworks were built for like, hey, does my website button work? Well, I'll write a test which is in code to click this button and make sure it works. So they're very deterministic frameworks, but the web is non deterministic, right? It's changing all the time. So people have been using deterministic frameworks to write non deterministic web automation workflows. And it means that their code breaks all the time. And that is like the biggest point of friction is that I write code to automate a website, the website changes, my code breaks. Now with AI, it is more possible to actually use an LLM to generate the code to control a website in real time. And that's been the biggest kind of like change in our, you know, we built a web automation framework called stagehand, which is built for kind of taking a like text based input and turning that into code that can control a website, trying to kind of help massage over the pain points these existing frameworks have. Now, you know, if you have a great framework to control a web browser, you're going to need good infrastructure to run the browsers. And that's where browser base comes in. And really I've firsthand experienced the pains of running headless browsers at scale. I've done that across two different companies now and there are just a lot of little gotchas that you run into because a browser isn't designed to run on a server. The Chrome team spends all their time focused on the millions and millions of laptops and desktop computers that Chrome gets installed to. They do not spend any proportional amount of their time to the headless browser use case. So that's on us. We have to be the headless browser experts and consider things like, hey, Chrome has a lot of memory requirements and you know that memory can leak if you don't properly shut it down. Hey, the Internet's a dangerous place. If your Chrome, running on your cloud, navigates to a website with a hacker and they download a virus, how are you isolating that within your cloud? So your cloud doesn't get hacked. These are the problems that our customers come to us to solve. But because they don't want to expose their cloud to the wild web, they want to have reliable browser infrastructure powered by a reliable browser automation framework. And that's where browser base plus stagehand kind of give builders the ultimate tools for automating the web.

Okay, so maybe let's imagine a world where like a GPT model never exists and you don't have that sort of base level of intelligence yet. Like what, what problems are you solving, you know, through sort of rules based systems or automation. And then how does recent innovations in scaling models and scaling of compute impact your ability to solve that? So where does AI, the recent advances in AI in particular scaled models with base levels of intelligence, you know, OpenAI just beat the ARC challenge. Like where does that come into play? What does that enable you to do?

Yeah, I mean there's a ton of progressive levels of automation here. Right. Like in the earlier years UiPath has been doing a ton of like RPA process automation. You can think about integration platforms like Plaid which are using headless browsers to log into bank accounts to retrieve data. Is that a form of automation? You know, it's adding integration layers on top. That's even pre LLMs. Then we have LLMs come in where we can now generate code from natural language which allows for us to do more on the fly browser automation, look at a page and then generate some code to control it. That's kind of a new change, but I think where there's still a lot to be desired for web agents, which are like AI agents that control the web is really in complex workflows. And we all know that if you built an AI agents step planning is a hard part of the process. And I think really complex step planning with high amounts of accuracy is solved more by scaled models. Taking time to think, doing chain of thought, reflecting on that, evaluating the plan before executing the plan. And you know, for us with stagehand, Stagehand is really more of atomic instructions where you can take like click the login button, extract the order from the page. It's not going to order you a pizza, it's here to help you reliably automate the web. And I think it benefits from scaled models offering better step planning because we aren't trying to solve the agentix step planning problem. Part of the problem for web agents or agents even generally, I believe that, you know, models like O3 will actually make agents that can do tool use and browser base as A tool much more possible. And we seek to offer a very, very reliable tool because your agent is only as good as the reliability of your tools. Let's say your agent has, you know, a plan it generates. If each tool only has a 75% chance of working, you're going to have a very unreliable agent. We seek to offer like a 100% reliable browser tool for AI agents so that you can build more reliability on top of that.

Well, I mean, I think actually, I mean this is an interesting discussion which is that let's say you have an LM which is non deterministic, it operates as your operating system. You, if it is, you know, you might, if you foresee a future where an agent is manipulating or interacting with other non determinant systems, you get this product rule where it's like you have 0.9%, let's say 0.9% probability of being accurate action times that by another set of actions, that's also 0.9. And so you get this sort of 0.9 times 0.9. And so you get a chain of certain non deterministic actions which to reduce your ability to be accurate, you need a tool that is sort of, if the element is going to, if the agent is going to call another tool, it better be performant and reliable. So that's, that's actually a very interesting idea. I haven't thought about the sort of chain of reliability. So like let's, let's break this down for people who are listening. Like, like what? So, like with your AI solution, like what, so what are the three key things it does? Like, well, let's say we're at a computer, we're using the tool. Just can you walk us through what is the problem I'm trying to solve and what are the three steps it allows you to do?

Yeah, when we think about how people use software, they're really trying to do three things. They're acting on a page, they're extracting information from a page, or they may be observing a page to gather more information. Stagehand, the open source web browsing framework that we've introduced, it kind of has three APIs. You can act, extract or observe. And the way that works with agents is that we offer those as three tools to an agent, where an agent which is maybe has a bunch of stuff, it has an email tool, it has, you know, a text message tool, and it has this web browser tool which is powered by Stagehand. And when it gets to a web page, it can say, I'll probably want to observe this page first and, and then see what actions are possible. Take some actions until I found the data I want and then extract that data in a type safe way so I can proceed forward my task. So imagine if you're ordering a pizza you want to. Your agent says hey, go to Google and take an action. Search for pizza websites and then once it searches, extract the pizza websites and extract the one that has the number one rankings in San Francisco. Okay, action. Go to that website and then order. Click on the pizza order button and then in action, make sure you got pepperoni to the cart. So by breaking this down into atomic actions, we're able to offer much higher reliability. Kind of talking earlier about like how you want to have more reliability each step more atomic steps are much more reliable than just a single tool call step which is get me a pizza, you know. So what we really think about is like how can we have stagehand really execute atomic reliable steps to help you build reliable web agents and prod. And we've seen success with that so far. But it all does come down to actions extractions and observations that happen on webpage to power that agentic loop for browsing the web.

Okay, so then what has been the most difficult part of building that system? So there's two frameworks you can use. We've gone from a world basically a few years ago where the emphasis was on fine tuning and training models to a new world of post training where we, we solve or find solutions through search through looking through the search space or we will do maybe some RLHF. So we'll use, you know, we'll use some, some, some, some word optimization. And I think, you know, there is a, I think Jim Fang from Nvidia said that give me a word function, I'll solve any problem or something like that. The idea is that with any sort of positive points or negative points you can get to a fully formed, a more performant solution. Like what's been the most difficult thing building your AI solution? Has it been on the training part gathering data? So some companies have access to just proprietary data. If you're a verticalized company who has maybe structured that into your contracts, you're able to train on that data. Maybe you can do steps to anonymize that data if you're a healthcare company to remove post identifiable information or is it on the search side, the post training side? Like where have you seen the most success in when you're building these AI solutions in terms of just like where you spend your focus?

Yeah, I think the hardest part of building reliable web agents in production is actually running the infrastructure. And that's why we've made stagehand our open source framework, completely free and open source. And we put all of our methods out there. You could go read the code and see, hey, how are they handling observations? How are they parsing the DOM in an efficient way? Because we think that the hard problem to be solved is actually on the infrastructural side. Like running a stateful distributed system of many headless browsers in the cloud is actually, you know, really, really hard and requires very specialized engineers. But we know that the power of open Source is important for achieving our mission of automating the web. And there will constantly be new papers and new methods for building reliable web agents. So we seek to keep Stagehand open source indefinitely because we know that we'll be taking inspiration from the best papers and the best research innovations out there. And the infrastructure is really, it's more something that we, you know, it's usage based pricing, you pay for what you use. And, and we think we deliver a lot of value via the abstraction of software complexity and the economy of scale that we can achieve when we build a great infrastructure company. So that's kind of like, you know, we, the value and the hard part is in the infrastructure, which is why we have to charge for it. And listen, the framework was hard to build too. You know, we've invested a lot of resources in it, but we know that doing that in the, for the good of the community is important because it empowers more interesting use cases that can actually automate the web in reliable ways.

Is there, you know, in the past year, is there a sort of a key innovation in AI, whether it's a paper you've read or you know, a piece of software that's come out that you really think is a game changer or just your jaw just dropped?

That's a good question. I mean, I think set of marks really changed a lot for web agents because what basically summary of this method was that, hey, we can actually present a UI to a VLM, like a vision model and just draw squares around all the things we care about and label the squares and say like which square is the appropriate square for completing an action, which may be like web automation. And I think that it's been interesting because we use set of marks or something similar to that within stagehand when you use the vision option because you can have vision enabled or disabled. And it's interesting to see how to me with such a naive solution, well, let's just draw some boxes on a UI and have a vision model, say which box is right, but also how hard it is when you have to draw many, many boxes that are layered. So we've seen a lot of derivatives from the original set of Mark's paper that has then resulted in interesting ways of marking up images for better VLM consumption or vision model consumption. I found that to be really interesting because it's one of those things where like, yeah, it seems like the naive solution, but it's actually quite good. So I think there's a lot of innovation that still needs to happen on web agents and web browsing for AI. We are so, so early on this and there's going to be a lot more to come for sure.

And do you foresee browser base for the future direction of the product? Do you want to go where do you want to go with it? Do you want to focus more on some of the agent side? Do you want to focus more on what you've done in the past? How do you think about future development or roadmap?

Yeah, maybe I can answer that question with a little bit more about me, because this is what inspires me to build the company I want to build. I started my career actually as an intern at Twilio during their ipo. So I saw the very end of a category defining infrastructure company. Twilio, it's a telephony API. They help send, you can send text messages, make phone calls, they even had a fax API at one point. So really kind of one of the early innovators in these API developer infrastructure companies. And to me that is the vision is to build a category defining infrastructure company because I believe that web browsing is a primitive piece of infrastructure required for AI. And what does that mean? Well, we really have to think about what are the primitives that we offer to our customer. Right now it's a web browser and. And then what are the things that we can kind of look upwards within the stack to support them better. So what I'm really interested about, like how can people run web agents on browser base? And Stagehand is kind of our first way to help out with that. What are the right frameworks that are taking off and how can browser base make sure we have integrations to support web browsing within those AI frameworks? We currently have that on Crew LangChain, all these Llama Index, all these popular Edge and frameworks. Browserbase has an integration there. So we really want to deepen our position as like, the browser tool for AI. And that means not only offering the best infrastructure, but also all the best integrations and abstractions to meet customers where they are as they're building things. But the space is changing so much. I think it's hard to really predict accurately where AI and what AI builders are going to be doing in the next year. So we're kind of staying nimble. And I'm not necessarily locking us into a roadmap of what we'll build in the next year, other than make the infrastructure as powerful as possible and make sure that it's something that people can depend on, because we talked about that probability chain earlier. If we're not highly reliable, it's going to be hard to build more agents on top of us. So reliability is just number one.

Okay, so then let's, you know, you know, if we, let's, let's just say that web agents are. They become mainstream and we are entering in this new world where, you know, websites are optimized for. Or, you know, maybe let's say, do you see a world where websites are optimized for bots versus humans? If you guys are successful, like, what does that future look like?

Yeah, I mean, I actually, like, maybe like the largest belief I have is that LLM, or it was called AI will be smarter than us, without a doubt. Like, and we will be the slower ones. So we're still going to have to have people using the Internet, and I think people are still going to have to click buttons. Even if AI gets really smart really fast, you're still going to people who don't use it in developing nations in the rest of the world who are still going to go to Amazon and order just the way they did before. So I think that we have to accept that since we're the lowest common denominator here, people are. You're still going to have to make websites for people. And to me, I'm lazy. I'll admit it. Why would I rebuild the entire Internet for AI when AI is going to be able to use the web just like people do? And you can kind of think about, like, how, like, robotics has been evolving.

I was. Exactly.

Yeah. Listen, I. That wasn't my analogy. Someone's way smarter than me said that. Right. As soon as you mention that.

I always think the exact same thing. The humanoid robots. It's better to train robots on humanoid form because we have all these videos on humanoids. Sorry.

It's funny that you said that, but I think it's also fair to criticize that and say, well Paul, that's not the most efficient way. And without a doubt we're going to see companies offering agent first integrations. I can expect all of the airline and travel industry will have agent first integrations. But you better believe that my barbershop, Joe's Barbershop in San Francisco will definitely not have an AI agent integration. And if my agent wants to book a haircut, it's going to have to go to the website or call them or do some sort of way to interact with these kind of real life establishments. And that's just a massive, massive long tail. The Internet is so, so big. There is going to need to be this primitive and we hope to power that. But parts of the Internet will change, especially in the top most popular way. So it's going to be a little bit of a hybrid. And you know, we still think there's a gigantic market here.

Yeah, yeah, okay. So we've, we've talked about the, the AI, like what you're doing in AI. We've talked about what does the future look like if you do succeed. I think that the most useful thing for a lot of builders is figuring out what to build and how to build it because they can go and read research papers themselves and you know, go through open source tooling. But I think for them it's, it's not just sort of how to build things, but figuring out just what to build in the first place. And I think one of the things that strikes me about your type of thinking is that you're very methodical. You know, I think a lot of people have a sort of a hack first mindset, which is great. But I think what really impressed me from your writings and you know, your appearances across sort of different platforms is the way you think through problems. So I guess could you just describe, you know, how you thought about this sort of being in a market where customers often will just do it themselves. We experienced this when building DSPY because a lot of people just said, why don't I just put my own language model in class? Why don't I just incorporate train of thought myself? If the dominant competitors build yourself, like how do you design browser base in a way to be both open source but also incentivize people to use the open source system as opposed to doing it themselves.

Yeah, I apologize to the listeners because I'm much better in written form. So this is my least good form. And I don't know, I think I just really try and think about all the things that can go Wrong and either accept them or come up with reasons why they're less likely to happen. And in my original memo for browser base, which is called an Internet browser for AI, I publish on my personal blog, at the bottom of the memo, I kind of laid out like every single reason that this thing couldn't work. And I either had like a counter position to it or a strategy or a justification. And I wanted to do that to make sure I could sleep at night and feel like, yeah, I believe in this thing despite all risks, because every company has risks. With Stagehand, why are we building open source? Well, we've just realized that like building a closed source means we'll miss out from the community that can be built around it. And like, I'm very happy to like let somebody run Stagehand on their own and try and operationalize it because I believe in the depth of complexity of the infrastructure because I've experienced that firsthand. So a lot of this kind of like act two of my journey in startups is very much focused around like what have I experienced firsthand and feel really comfortable in. I've operationalized headless browsers at scale before and I've wasted a lot of my time dealing with it. At my last startup, I just have leaned towards that and I think that is one of the benefits of being a second time founder. I told a lot of people after my first company I would never do this again. I was done. So when I kind of came back and said, well, maybe, maybe I'll do one more, it was really a lot of myself trying to talk me out of it because it's not worth building a company unless you have extremely high conviction. Yeah, that's, that's really the, the most important thing with building companies, regardless of anything else.

So maybe, so to wrap it up, I would, maybe we can break this down into lessons of the past, lessons of the present and the future. So for lessons of the past, like what have been the. What's the one key thing that you learned the second time around? Or, or did you know you're, you're. However times you started coming, I know you had the part company you sold, but any other sort of experience that you had in the past, like what did you learn this time and what did you sort of focus in on this time?

Yeah, I think from the past it is so important to do things for the right reasons. And I'm doing browser based because this is truly the company I believe in the most right now. I think with my first startup I kind of did it because starting a startup seemed like the next logical step in my journey, but I didn't really deeply believe in the idea as much as I do with browser based. So I think really building that kind of like deep, deep conviction in your space and understanding it deeply is a prior step for building a great company now. I also did a lot of upfront thinking. I wrote a whole 3,000 word memo about browser before I even raised a single dollar. And that memo has ended up being the roadmap that I followed for the first year that's allowed us to get to a million dollars of run rate and hundreds of customers and a team of 15 in under 12 months. And I think that execution comes from the depth of thought beforehand of building the company and then finally it's being able to kind of learn from similar companies that have existed before you. With Browserbase, my deepest inspirations are Vercel, Twilio, Stripe, aws, Datadog, some of these amazing infrastructure and developer tools. Companies that are operating in very separate spaces but have overcome similar challenges that Browser Base will have to overcome in the future. When Twilio launched, they're talking about robocalls. When browserbase launched, they're talking about captchas and bots on the web. Very similar regulatory problems that we'll both have to mature with as these spaces mature. But in the end I'm super optimistic because others have done it and the goal is always to be building a category defining company. And I believe that browser is a category defining company in browser infrastructure.

What did you learn from Vercel? I'm curious.

Yeah, I think Vercel has really done a great job of balancing open source and infrastructure. They have this kind of idea of framework defined infrastructure. It's an amazing blog post. And you know, Vercel is to browser base as next JS is to stagehand, you know, build the framework and then build the infrastructure that's perfect for that framework. And I think they've done an amazing job of balancing an incredible open source product as well as one of the best in class hosting products for that infrastructure.

I mean, I mean this is what made Databricks and Apache, Kafka and Confluent perfect. They created the Apache Cockpit software and guess what we're the best at creating. Confluent is the best at creating the platform that it's built on or it's run on. So I like that way of sort of articulating it, I guess. And then that sort of lessons of the past, I mean, I guess lessons of the Present. I mean there's been a lot of news from the O3 model to deep seq AI you trained. Mixture of experts model recently that basically was pretty cheap, was really highly performant. I mean there's just been, you know, metas, meta recently their llama models have you know, significant improvements in performance almost to you know, private models. Like is there something that sort of catches your eye or you were, you know, was. You think there's broader implications for the AI community.

I think that as an infrastructure founder I'm really happy to see healthy competition at the labs level. You know, to see if Anthropic, OpenAI, Lama, deepseek, Gemini, like just a lot of really good, you know, variety and choice for AI builders, which I think is really important. I am the biggest believer in OpenAI and I would buy OpenAI stock today if I could. However, I think it would be quite bad for the consumer if only OpenAI was a successful lab and it would be impossible for a startup like Browsebase to exist because as consumers move from one AI model to the other they want infrastructure that comes with them and browser base can exist in that category because there is healthy competition at the model layer. And I think like the power of open source here has been really interesting. Like you can run Deep SEQ on your own and that's like, that's pretty, that's pretty exciting. You know, I'm really happy to see that continue to happen. In terms of innovations that I'm surprised about, I think we're going to start to see labs move up the stack more and we're going to see more agent products from labs and once again I think if, as long as there's a lot of choice for customers, I think it's exciting. You know we are, we're really early in AI and I think it's not a, not currently a zero sum market at all. Maybe in the longer term it will be but more, you know, tools for people to build powerful pieces of software with AI no matter the size. That's going to just be a huge unlock in productivity and hopefully gdp. So bullish on global GDP with the power of AI over the next decade or so.

Yeah, I think, you know, one of the things that I was thinking about is just like coming from the browser perspective, is there something that you, is there a perspective on AI that you think that you have that most people wouldn't have? I'll give you an example. So if you, if you work on in healthcare automation, you would think that some of these models are Nice, but they're not super helpful because you still can't have accurate diagnostics with a medical chatbot or if you're in finance, it's nice to be able to tell me my revenue. But you know, can you. And you can just spit out some risks, but can you sort of thoughtfully go through and analyze. These are the key risks that really matter. However, if you are in the coding space, you have completely different perspective. What your perspective would be is that with instruction fine tuning from some new data sets like instruction, if you were solving coding problems or math problems, you look at star coder or capital letters math and all these other data sets and advances in instruction fine tuning, you'd say no, this is groundbreaking. We'll be able to automate many of the systems that engineers build. Do you think that there's something unique to browsers that gives you a unique perspective on the problem of AI?

Well, I think the browser automation problem is an extension of coding because you're writing code to control a browser. So browser automation does benefit from the advantages that fine tuning a coding talented model has. You can write tests that control a browser that pass or fail and evaluate a browser with code. It's much harder for a financial analyst or a healthcare person because I think those evals are just harder and more human driven. There is quite a lack in web agent evals and Stagehand has quite a few within our repository just so we can measure the improvement in progress. So I kind of like, once again I think that web browsing is a primitive but, but it is downstream of code execution. And I think if you asked me for a hot take, I think my hot take on this state of AI right now is that when people think of things like computer use, oftentimes these kind of computer use models, like the one for Anthropic, they're controlling a whole desktop. I don't think it's going to be desktop control that's needed. I think it's just going to be the browser. The majority of computer use demos I see are with a web browser. And sure it's necessary to have LLM LLMs control a whole computer. If you're like automating super legacy software, like you know, maybe legacy ERPs or electronic health record software like EPIC is desktop only in some capacities. However, the majority of software is online and we've made, we actually like have timed it really well where we have made a lot of the shift from installed software on your desktop to software running on the web somewhere. And I do think the majority of like productive work that AI will automate will happen in a web browser. And that's why I'm like super excited about browser based because we, we just can have a lot of market share to capture. We, we ran the numbers on this in like a slide deck and it was like north of a trillion dollars. So I'm, I'm feeling very, not zero sum right now. Like it feels like it's a lot of, a lot of stuff that can be building and it's cool to see everyone kind of innovating in the web web agent space.

Okay, so that's a really hot take. And I, and what I love about that is that if you think about, you know, maybe a few years ago, the on premise versus SAS debate whether you should build for on premise or SaaS, people would say well there's such a huge market for on premise. And then there was just these growing pains in the 2010s and 2020 where you know, Adobe had to basically they, their stock was crashing because they had to pivot from on premises. But once they did that they were just a behemoth. I mean you have your workflows that drive every single company on the planet basically you have with the SaaS offerings to recurring revenues or you can do, you can do add ons easily and immediately. And so those who had bet on the SaaS shift, although there was such a big piece in the on premise market, those, those basically won. And so what you're saying with sort of the, the browser versus sort of desktop know building for either one, it's almost like you're sort of building for the future and you're betting on that future as opposed to sort of building for the past. So that's it. I love that take.

Both are necessary too. Listen, like once again like we should still let AI control a computer. I just don't, I think that's overkill and it's going to be very, very expensive to run a full desktop environment that's secure and can be interacted with via AI versus like I had this browser as a factor of 10 less expensive to run. And if it's as valuable as the full desktop environment, I just think the desktop environment for AI won't be the default. It will probably be like a browser among some other tools that would be very important like access to your email or access to your phone, you know.

Yeah, no, I mean I think especially because I have to make the point but in sort of security prone environments like if you think about you maybe a healthcare system, they're not going to want to give you access to the entire desktop. You probably maybe, you know, in that case would want to just run it in your just focus on the browser and give you access to that as opposed to everything else. So yeah, that is a really brilliant take and I think with that, you know, Paul, I just want to thank you for propping on and I really appreciate you sharing your insights.

Cool man. Yeah, thanks Thomas. I had a great time. And we are hiring at browser base. So if any listening that wants to come join us on this crazy mission browserbase.com Cruise sweet. Thanks Paul.