AI Engineer World's Fair 2025
The Web Browser Is All You Need
Read the talk
The Web Browser Is All You Need
When a service has no API or MCP server, browser tools give agents a way in. The engineering challenge is choosing how they perceive pages, control actions and handle unexpected results.
From a talk by Paul Klein IV
Before you start: Familiarity with language-model tool calls, basic browser automation and MCP will help; no prior Browserbase experience is required.
What should an agent do when there is no integration?
What happens when a user asks an agent to do something on a website that has no dedicated integration? Before the agent can navigate, someone has to run its browser. Paul Klein IV, founder of Browserbase, opens with that infrastructure problem: maintaining headless browsers yourself can be fragile. He describes Browserbase as supporting thousands of headless browsers in the cloud, with browser automation exposed through its MCP server. The focus here is what those browsers make possible inside agent applications.
Consider the DMV or a local barber. The DMV is unlikely to offer an MCP server soon, and Klein’s barber has better things to do than build a GraphQL API for appointment booking. Both already have an interface people know how to use: a website. The browser connects agents to services that were built for humans.
The useful opportunities include less glamorous work than booking flights or picking restaurants. Klein’s example is filing Delaware franchise tax: an administrative task founders need done, whether or not its website attracts much developer attention. When MCP, A2A, OpenAPI or a first-party integration is unavailable, the agent can still use the website. The browser is an integration of last resort, a route to take when a suitable primary integration does not exist.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From page context to browser actions
A web agent needs to turn information about the current page into an action. One approach gives a model the page’s DOM and asks it to generate browser-control code. The ecosystem Klein describes includes WebVoyager, Adept’s Fuyu models, OpenAI’s Operator, Proxy by Convergence and H Company. These systems explore how models can operate interfaces rather than merely describe them.
WebVoyager is an early example of the visual approach: take a screenshot, use prompting to reason about the next action, then select a target on the page. Visual targets can be represented by coordinates or annotated labels. Text-based agents instead consume HTML and identify targets through structures such as XPath or Playwright locators.
| Approach | Page context | Action target |
|---|---|---|
| Vision | Screenshot, sometimes annotated | Coordinate or marked region |
| Text | HTML or DOM-derived text | XPath or Playwright locator |
Klein associates text-based control with repeatability and visual control with handling some complex pages, but neither is a universal winner. The right representation depends on the website being automated.
With set-of-mark prompting, the screenshot contains labeled boxes over possible targets. Instead of requiring the model to describe a button’s position precisely, the system can accept a selection such as box 25 and map it to the marked target. On the text side, an accessibility tree offers a more compact semantic representation of page elements, stripping away much of the surrounding div and class scaffolding. It exposes useful structure without needing to reproduce the page’s exact visual layout.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Learning a path, then deciding who controls it
Computer-use models move beyond asking general image models to interpret individual screenshots. Training on web trajectories gives a model sequences of interactions: what it saw, what it did and what page followed. Reinforcement learning can use those sequences to teach decisions across multiple pages. The problem becomes choosing a path through a website, not merely choosing the right button on the current screen.
That capability raises a separate design choice: how much of the path should the model control? In Klein’s terminology, a web agent turns one prompt into many actions. Ask Operator to file Delaware franchise tax and it must decide how to reach the goal. Repeating the prompt may produce a different route. Its ability to keep looking for a way forward comes with nondeterminism because the agent owns the reasoning between steps.
A browser tool accepts a narrower instruction, such as clicking the sign-in button. Purchasing an item is a sequence of steps; clicking sign-in is one bounded action within a sequence. In Playwright, the latter can be expressed directly:
typescript
import type { Page } from "playwright";
export async function clickSignIn(page: Page): Promise<void> {
await page.getByRole("button", { name: "Sign in", exact: true }).click();
}
The surrounding application still decides when to call this function and what comes next. That is the control boundary: selecting a target for one action versus planning the whole task.
Stagehand provides the browser-tool approach Klein recommends for workflows whose high-level steps are already known. The application supplies those steps, and the tool translates them into browser actions. If the incoming request is open-ended and the steps cannot be anticipated, a more general web agent may fit better. The Stagehand and MCP repositories linked here document current APIs; their installation examples and defaults should not be read as the historical implementation used in the recording.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Application-specific tools and reusable primitives
The same distinction between a specific operation and a general capability appears in MCP server design. A vertical server exposes actions for an application or domain. A horizontal server exposes primitives that can work across applications.
| Server scope | Example | Exposed operation |
|---|---|---|
| Vertical | Linear | Create or assign a ticket |
| Horizontal | Browser | Click a button on a page |
The browser’s page can change while the primitive remains useful. That gives one server broad reach across websites.
Broad reach does not make direct integrations obsolete. If Salesforce has an appropriate MCP server, use that rather than driving its UI unnecessarily. A bespoke enterprise CRM without an MCP server is a different case: browser tools can operate its existing interface without first reverse engineering its APIs.
Approval is another part of the integration boundary. Klein expects security teams to scrutinize dynamic tool discovery—the prospect of an agent attaching itself to an effectively unlimited collection of MCP servers. A horizontal server can reduce the number of servers an organization has to onboard. That simplifies one part of approval; it does not grant blanket permission for every website or action reachable through the browser.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Evaluate the task and inspect the run
Klein is skeptical of public benchmarks, particularly rankings published by the company ranking its own system. The practical response is to build evaluations around the actual websites and tasks the application needs to handle. He names Braintrust as supporting Browserbase’s evaluations: the useful question is which model works best for a particular web task or site, rather than which model leads a general leaderboard.
Evaluation tells you whether the outcome was right. Observability helps explain how it happened. If an agent is supposed to buy an Xbox but buys AirPods, the final result alone cannot identify the mistake. You need the prompts, screenshots, browser actions and navigation history to reconstruct where the run diverged. Klein describes Browserbase as bundling session recordings and logs for this purpose.
These requirements qualify the browser’s role as the default MCP server for the rest of the internet. When a suitable API or dedicated integration is absent, browser access supplies a path forward—but that path still needs evaluation and a record of what the agent actually did.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Finding an adoptable dog through Cursor
The live demonstration puts a reasoning agent above browser primitives in Cursor. The request is to create a browser session, navigate to sfspca.org, close any modals, find a dog available for adoption in San Francisco and return its URL. This is a goal with several intermediate actions, not a single predetermined click.
The visible sequence makes the division of responsibility concrete:
- Cursor calls the browser session tool to create a session.
- The agent issues individual tool calls to interact with the page.
- Klein opens the browser session beside the agent interface so its progress is visible.
- The agent works on closing a modal and navigating the site.
- The reasoning model chooses among the exposed tools and returns a dog’s URL.
The browser tools supply the actions; the model determines which action to request next.
Opening the result brings another interruption: a fundraising modal complicates the reveal. This is exactly the kind of page behavior that makes fixed expectations brittle. An agent needs to react to what is currently on the page, including overlays that were absent or already dismissed earlier. After the interruption, the visible adoption page shows two dog photographs and the heading “Meet Zoe” alongside the agent chat. The demonstration reaches an adoption profile; it does not complete an adoption transaction.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The opportunity in ordinary operations
The broader application is bringing capable models to mundane work in organizations that are not AI-native. Klein describes a 55-year-old dairy trucking company that hired its first engineer that year and used Browserbase to automate a painful operational workflow. The example matters because the value comes from reaching an existing business process, not from requiring the business to rebuild itself around a new agent interface.
Klein closes the prepared portion by inviting the audience to try the available browser MCP server through a signup QR code. The adoption step is to give an agent browser access and apply it to a workflow that otherwise lacks an integration.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Bring the model; separate the infrastructure
The first substantive audience question asks whether navigation requires a particular model. Browserbase’s boundary is model-agnostic: it supplies headless browser infrastructure, an MCP server and frameworks such as Stagehand. The application brings the model that decides how to use them.
Model selection also depends on operational requirements. Klein uses HIPAA and zero data retention as examples of considerations that may affect provider choice. Those are requirements to assess, not compliance guarantees conferred by a model name.
For the demonstration, he confirms that he used Cursor’s agent mode, then tentatively identifies the selected model as Claude Sonnet 4 after correcting himself. The key architectural point does not depend on that identification: Cursor’s agent supplied the reasoning while the browser infrastructure exposed the session and actions.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Let people inspect and intervene
A financial-data question makes the distinction between visibility and control more precise. An advisor may need to review information before it enters a financial planning tool. Users may also need a clear account of what an agent did, even if they did not participate in the run.
Klein describes two capabilities:
- Session recordings: Embed a completed run in a user-facing application so someone can inspect what happened. Recording can be enabled or disabled according to data sensitivity.
- Live View: Embed the running browser through an
iframe, allowing a person to watch and also click or type in the session.
Live View supports a browser copilot experience in which a human can take part in the interaction. These capabilities provide visibility and intervention; a mandatory approval gate before submitting financial data remains an application-level requirement.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Authentication and the limits of browser access
CAPTCHAs expose another boundary: a website may resist automation even when the agent has a useful task. Klein distinguishes beneficial automation from malicious bots and describes Browserbase’s CAPTCHA-solving and proxy capabilities. His example includes customers encountering their own CAPTCHAs when their security operations process has not provided a suitable automation route.
His longer-term hope is agent authentication. Klein says most CAPTCHAs Browserbase encounters arise during login. If a site can identify both the agent and the person it represents, it has a stronger basis for deciding what that agent may do. He presents authentication as a potential route away from CAPTCHA solving, not as an established solution that every website already supports.
The final question asks whether automated navigation itself increases the likelihood of a challenge. Klein answers that detection often depends on behavior. He points to LinkedIn’s monitoring of actions per minute as an example: browser stealth features cannot guarantee access when an agent’s behavior crosses a site’s limits.
His operational advice is to be a responsible participant on the web: respect robots.txt, act carefully and recognize that infrastructure cannot prevent every block. Browser access does not erase a website’s constraints or the obligations of the person operating the agent. Klein ends by rejecting unlawful or unethical use of the platform.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Official browser automation MCP server, with current setup instructions and browser tools built on Stagehand.
Browser automation SDK combining browser APIs with natural-language actions, observation and extraction.
Research on a multimodal agent that completes tasks by interacting with real websites, with a task benchmark and evaluation protocol.
Documentation for embedding a running browser session and allowing a person to watch or interact with it.
Further reading
Original research on labeling image regions to help multimodal models refer to visual targets.
Paul Klein's June 2025 account of Browserbase's infrastructure thesis and the value of reusable automation primitives.
Updates since the talk
- OpenAI API data controlsDocumentation
Current documentation on retention controls and endpoint-specific data handling restrictions.
Read the complete timestamped transcript
- 0:00
[on hold music] Hey, everybody.
- 0:16
I'm Paul. I'm the founder of Browserbase, and I am obsessed with browsers, specifically one type of browsers, headless browsers. And I'm here to talk about how the browser is all you need.
- 0:27
It's not attention, it's not MCP, it's the browser, or specifically the Browser MCP Server is all you need. And I'm gonna try and keep it light on slides. We only have a hundred to get through in twenty minutes, but you and I, we can do it together.
- 0:40
Uh, just stay with me for twenty more minutes, okay? So first of all, a little about Browserbase. This is what it looks like to build your own browser infrastructure.
- 0:47
It's messy and it breaks all the time. With Browserbase, we l- can let you run thousands of headless browsers in the cloud for easy to control. And you might be wondering, like, "Well, why do I need a browser?"
- 0:57
Well, every AI agent needs a web browser. That's the point of this whole talk. So of course, you know, I can talk to you about how to scale browser infrastructure and all that interesting stuff, but that's too pluggy.
- 1:06
We'll save that for later. But I do know quite a bit about, you know, how customers are using browsers as part of their AI agent applications, specifically with MCP.
- 1:15
We have the most popular browser automation MCP server out there. And the reason why people choose Browserbase to run their headless browser MCP is that we make it really nice and, and really can help you scale your infrastructure.
- 1:26
And when you think about, why do I need a browser? Well, I like to think about two things. You have AI agents and the legacy internet. You know, the DMV is not gonna have an MCP server anytime soon.
- 1:37
My barber shop is not gonna open a GraphQL API for me to schedule a haircut. As much as I keep begging John to do it, uh, he's got better things to do.
- 1:47
So if we want AI agents to interact with the rest of the legacy internet, they need a bridge. And I really do believe that the browser is that bridge between AI and the rest of the internet.
- 1:56
And this is the unsexy internet, I might add. It's, it's the internet that's not gonna get a lot of attention. I've seen countless flight bookers and countless restaurant pickers, and I have not seen anyone do the thing I need, which is a Delaware franchise tax filing agent.
- 2:10
Any founders in the room have done that before? Not super fun.
- 2:14
And I think people use a lot of acronyms these days [laughs]. You know, you have MCP, you have A2A, OpenAPI. But if those aren't available, you can just do what could be considered the, the dumb thing.
- 2:23
You could just use a website. And websites are out there. There are plenty of them. There are billions of websites. And when your user is gonna prompt your agent to do something, you might not always have a first-party integration available.
- 2:34
That's where a browser is kind of the integration of last resort, the path that you can take your agent down if you don't have something, a primary integration already built in.
- 2:43
And you might be wondering, "Well, okay, Paul, cool, I get it. You're beating this thing dead. Uh, AI agency and the browser, well, how do they control it?" Well, you can think about web agents and browser tools.
- 2:52
First of all, you know, what is a web agent? Uh, I wanna keep this technical, as this is a bunch of AI architects, right? You know, web agents, we've, we've heard about them for a long time, take a model and then have it generate some code to control a browser by generally parsing the DOM of the page,
- 3:06
the HTML and the CSS. WebVoyager was early here. Adept did a lot of really cool stuff here with their Fuji models. OpenAI had Operator, Proxy by Convergence, now Salesforce, Hcompany.
- 3:16
Everyone was kinda doing a lot of this stuff last year. And we really got to see a lot of web agents in production, uh, but they were still early days, you know?
- 3:23
And WebVoyager was first, you know. It was taking screenshots of a page. It was using chain-of-thought prompting, and then from that, it was saying, "Click the button at this coordinate."
- 3:32
Uh, sometimes they're doing these things, we'll talk about in a second, called, um, labeling on top of the page. But I think it's pretty cool 'cause we haven't changed that much from this.
- 3:39
There's really are two different types of web agents. There's vision-driven agents. These are ones that predominantly use screenshots as context for the model. They might do some marking up of the screenshot to indicate what box to click on.
- 3:50
Um, or there's text-based web agents who predominantly use HTML as the context of the model. Um, both have different approaches, pros and cons. Text web agents use XPaths and Playwright code.
- 3:59
Some may argue it's more repeatable. Vision models can be more accurate on more complex pages. Uh, there's trade-offs here, and they're, they're, you know, it really d- does depend on what website you're trying to automate.
- 4:08
Here's an example on a vision agent using set-of-marks prompting. You can see these little boxes here where you're marking up what you should click, and the agent will, or the model will turn and say, "Click the box label twenty-five."
- 4:20
That's gonna help you out. And on the DOM-based agent side, there's also ideas of, like, taking HTML, and how can we transform HTML to make it more reliable for web agents?
- 4:29
So the accessibility tree is something that's built into every page, uh, in a lot of applications, where you can take a different structure of the same information and condense it down where you get the same layout, but without all the extra div tags and classes.
- 4:42
So we have vision-based agents and DOM-based agents. And there's now also computer use models, which are kind of like the next step here, where, well, let's train a whole model on this stuff.
- 4:52
Previously, we were just using, like, the stock image models, but now we can train that model on these things called web trajectories. And I won't go too much into this.
- 5:00
There's a lot of great papers out here. I'd recommend this paper rec- linked down here about, like, web trajectories and how you can generate them to do RL, to teach models how to not just pick the right button on the page, but how to reason across multiple pages about the right path to take.
- 5:12
But all in all, there's just a lot of innovation here happening on teaching AI how to browse the web, and this stuff is getting good. It is working, and you can use it today to help add some sort of extra functionality to your applications if you're making the right choices.
- 5:28
And I, I will kind of add, you kind of want to think about if you want a web agent or a browser tool. And you may have never heard of, like, what the difference is here, but there is a difference in my head.
- 5:36
Like, a web agent is kind of like one prompt to many actions. I think OpenAI's Operator is a good example of this. You say, "Hey, Operator, go, you know, file my Delaware franchise tax."
- 5:46
You give it some, some prompt, it's gonna go take many actions, and if you give it that same prompt twice, it might take two different paths to get the task done.
- 5:54
Web agents are good. They're like little cockroaches. You just-- They're just gonna keep trying to find a way to complete your task. Um, but they're a little bit more nondeterministic because the reasoning is in their control.
- 6:03
Whereas like browser tools are like one action or one prompt. You say click the sign-in button, that thing's gonna click the sign-in button. You ask it to purchase the Amazon, uh, item that you want, that's a series of multiple steps that really might be more suited for a web agent.
- 6:16
So we have a framework called Stagehand that we think is the best browser tool. Um, and it really does depend on what you want. If you know what your workflow's gonna be with some high-level steps, you can actually use a browser tool to take those steps and translate them into reliable web, web actions.
- 6:30
If you don't know what you're gonna do, if you, if you don't know what the prompt will be, a web agent more generically might be the right fit. But I do believe your AI agent needs a browser tool.
- 6:39
Another thing I'll add is, like, you may wanna think about with MCP, what types of servers are you integrating? So a vertical MCP server with something like Linear, where it's going to give you options to control specific things on a specific task, like create a Linear ticket, create, uh, assign someone to the Linear ticket.
- 6:55
Whereas like a horizontal MCP server, you're going to have some sort of primitive that can do many things. You know, for browsing, we view it as a horizontal MCP server.
- 7:04
You're exposing primitives like click a button on a page. Now, that page may change. There might be many different pages. But when you have a horizontal MCP server, with one server, you have the opportunity to automate the whole web.
- 7:16
And these, like, you know, you will see vertical MCP servers, which are more direct tool calls, as, like, an important part of your agent. I'm not saying we replace those with browsers.
- 7:24
If you are interacting with Salesforce, you probably should just use the Salesforce MCP. You don't need a browser there. But if you're interacting with custom bespoke MCP built by large enterprise that doesn't-- Uh, sorry, a custom bespoke CRM built by a large enterprise that doesn't have an MCP server, you don't have to go reverse engineer APIs.
- 7:42
You can use a browser tool and a browser MCP server to go out and automate that. Is that making sense for everybody so far?
- 7:50
Yeah.
- 7:50
Okay. Thank you, guys. I'm hearing some heads nodding. Okay, a few other notes on like MCP and, and all these, like, concerns I have here. Um, I, I do think that MCPs are gonna have to pass compliance, and dynamic tool discovery, this idea that you can plug in your agent to an infinite toolbox of MCP servers, is
- 8:08
going to be hard for CISOs to get on board with, right? You're gonna wanna be able to select which MCP servers make the most sense. And with a browser or a horizontal MCP server, this could be browse, it could be email, it could be anything, uh, you really only have to onboard one MCP server as opposed to
- 8:23
do an MCP server for each individual integration. Secondly, like, yeah, that's the most important way. Secondly, like evals. Most, most benchmarks are fake news. Just wanna let you know.
- 8:33
Especially when the company putting out the benchmark is the one that's also ranking themselves. So I would be very, uh, critical of public benchmarks of any web agent you see out there or any model you see out there.
- 8:43
You really need your own evals. My friend Ankur, who runs BrainTrust, he supports our evals. I really like them 'cause then we're able to say, "Hey, actually, for our web agent we're building, which models are the best model for this web task or for this certain website?"
- 8:57
And you can get really intelligent and honest about what you actually need to do to automate the web and which model is the right choice for you. Finally, I think you need observability.
- 9:07
If your AI agent is controlling a browser, you need to be able to see what happens in the browser. Browserbase bundles this out of the box so you can build your own, you know, taking screenshots, recording history, recording actions, making sure that you know exactly where your agent went to and why is important because let's say your
- 9:24
agent is gonna go buy an Xbox and it buys you AirPods, you wanna understand what prompts went into that, what page paths it took, and really break that down.
- 9:31
At Browserbase, we include this in every browser. Its sessions are recorded, logs are available. It really makes it quite easy. Okay, so that was, like, a lot on browsers, and I got a ton of time left, but I'll end with this one point.
- 9:44
The browser is the default MCP server for the rest of the internet. If you need to integrate with something, whether that's with MCP, an API, and there's not something available, you should really consider including a browser, 'cause a browser is all you need.
- 9:57
And since I'm doing so well on time, I'm gonna do some live coding, 'cause I feel like there hasn't been enough live coding in this room today. So I'm breaking it out, and let's pull up Cursor and bring it over here, and it's gonna be so hard to see, but let me try.
- 10:10
Oop. Give me this, please. Love. All right. So we have Cursor right here in all of my screenshots. I'm a screenshot hoarder. I apologize. So I pulled up the Cursor MCP server.
- 10:26
Uh, sorry, the Cursor controller, and I've written a prompt. It says, "Create a new browser session, navigate to sfpca.org, close out any modals, and find a dog for adoption in San Francisco.
- 10:36
Return the URL." We'll click enter, and I may have to jump to a browser here really quickly. So you can see it's calling the browser session tool. It's generated a browser session right here.
- 10:48
And you can look how it's actually making these individual tool calls on the page. If I pull up the session at the same time and scoot this tab over there,
- 10:57
we can see in parallel the browser is loading. And as our MCP server is making these tool calls, right now it's trying to close out the modal, it's navigating the website.
- 11:07
And thinking about how, like, there's a reasoning model here that's deciding, "What should I do? What should I call?" Uh, these tools all are exposed. And now, if we go here, we have been given a dog with this URL, and one of you will be going home with a lucky dog today.
- 11:25
All right? Can I get a drum roll, please? [drum roll]
- 11:28
All right. That was so half-hearted, but thank you. Uh, give me a dog. Yay. [clapping] There we go. Oh, they really want that 200K match. Uh, this campaign's been running for a while.
- 11:40
I feel kind of bad offensive today. Uh, but that's a good example of what will happen when you're building these web automations. Sometimes modals will pop up. Sometimes things you aren't expecting may happen on the webpage.
- 11:50
You need to have an AI agent driving the page with primitives so it can react to any sort of weird changes that happen. And hopefully, you can integrate with things that aren't gonna be AI native.
- 12:00
To me, the most important problems to solve in AI right now ... are the unimportant boring problems, and they're gonna require, you know, intelligent, you know, engineering to both bring the best models to the unsexy problems.
- 12:15
And when we talk to customers at Browserbase, they're not just Perplexity or Clay or, uh, Camure. They're also a [REDACTED:age] dairy trucking company who has never hired an engineer in their 55 [REDACTED:age] until this year, and the first thing they did was use Browserbase to automate a really painful ops workload, workflow.
- 12:32
So if I can somehow pull my slides back, I don't know if I'll be able to. I might just have to move this bad boy over here, and we'll just go for it live.
- 12:39
Uh, yeah, screw it. Well, let's do this. Okay, so if you do wanna try a browser MCP server, it's available today. You can actually go ahead and scan this QR code and sign up or use this to try adding some sort of automation.
- 12:52
It's really easy. If I can use it, you can use it, I promise. I'm gonna pause. I'm seeing some photos being taken. Amazing. Great. And, and finally, you know, if you are looking to, uh, join a company that's growing quickly, Browserbase's been around for a year and a half.
- 13:05
We're 30 people. We're backed by some really great investors, and we would love for you to come build the future of automation with us. All right. That's all my time. [clapping]
- 13:13
I might have a minute for some questions or two, if that's okay. Yeah, if there's any questions, uh, happy to, to take them. Otherwise, uh, thanks so much, everybody.
- 13:21
Any questions?
- 13:22
Do you use specific model-
- 13:24
Hold on. Maybe we do hands, if that's okay. [laughs] Uh, we'll go, uh, front and then back. Yeah, go ahead.
- 13:28
Do, do you use specific model, uh, do you use specific model for navigation, like it's your own model or, uh, like, uh, when the browser is-
- 13:37
Mm-hmm
- 13:37
... scrolled and so on and so on-
- 13:38
Yeah
- 13:39
... or you utilize different models?
- 13:40
We are model agnostic, so Browserbase is just the infrastructure for running headless browsers as well as the frameworks and tools. We have an MCP server. We have a framework called Stagehand, which is like Playwright but better.
- 13:50
Um, you bring your own model, and you may wanna choose different models based on your conditions. If you are doing HIPAA and you need like zero data retention on OpenAI, you might wanna use one of their models.
- 13:59
So you bring the model, we bring the infra.
- 14:01
So with that example you just showed-
- 14:03
Hold on. I'm sorry. I'm just gonna... Go ahead. Say it again. You can just... We just skip it.
- 14:08
The example you just did.
- 14:08
Yeah, the example we just did-
- 14:09
Cursor agent versus-
- 14:09
I used, uh, yeah, the cursor agent mode. I think it's, uh, 4.0 max. Or sorry, Claw- it's probably 4.0 Sonnet right now is what I have baked in there.
- 14:18
Go ahead. [laughs]
- 14:19
How do you-
- 14:19
Love the questions. Keep 'em coming.
- 14:21
How do you guys manage, uh, anything that requires human in the loop? Because we have to deal with lot of legacy infrastructure, but it's financial data, so an advisor wants to review it before it goes into a financial planning tool or something like that.
- 14:34
How do you bring in some sort of human in the loop interaction or to a browser, A? And B, if, um, uh, uh, because this is financial data, we also want to, um, give users a clear view of what the agent has done.
- 14:49
Is there a way we can send that information even if they couldn't interfere in the process?
- 14:54
Yeah. I, I was just gonna hop over the slide. So not only do we have recordings available, and they can be turned on or off depending on data sensitivity, you can embed these recordings into a user-facing application and show your user what happened.
- 15:04
We also have this feature called the Live View, where you can embed an iframe and show exactly what's happening in the browser. And better yet, if someone wants to, a human wants to come in and click and type in on the Live View, they can do that as well.
- 15:16
So it's not just for browser automation. It can be a browser co-pilot, and it's an iframe, so it's a browser inside a browser, which is kind of fun to see.
- 15:24
Yeah. Thank you. I think there's a question over here. Yeah, no, we'll go to you guys next.
- 15:28
Um, so my question is kind of two parts. One, have you dealt with CAPTCHAs at all, and is there-- And do you see like future websites kind of using similar strategies to defend against like automation tools, that it's only users can use their website?
- 15:42
Yeah, CAPTCHAs. So for the longest time, there's never been like good bots on the internet, and CAPTCHAs were built to stop all bots. But now there are good bots and bad bots.
- 15:51
And at Browserbase, we do offer CAPTCHA solving and proxies as something that's needed to browse the web. We have customers that use CAPTCHA solving against their own CAPTCHAs because they can't get their SecOps team to like get through a, a good way to bypass that.
- 16:04
However, I think longer term, you know, my friend Michael at WorkOS was just talking about CAPTCHAs. Agent authentication is the path to avoid CAPTCHAs. Most CAPTCHAs we see at Browserbase are when someone's logging in.
- 16:14
Once you log in, you know who the agent is. You know who they're acting on behalf of. And I'm really hopeful that solving CAPTCHAs at Browserbase is a short-term solution as we figure out how to do agent authentication on the internet longer term.
- 16:25
But down to talk about that afterwards. Come find me.
- 16:28
I got, I got one minute 42 seconds. I'm gonna... I'm holding us to it. Yeah. [laughs]
- 16:32
Uh, so during web navigation, like I think my question was related to CAPTCHAs, but during web navigation, if, does the browser actually detect it's a robot that's doing the navigation, and does it increase CAPTCHA coming up?
- 16:45
Yeah, I, I think that the way that CAPTCHA detection works is often based on your behavior. And what we advise our customers is like, "Listen, you know, even though we provide the best stealth browsing features necessary, in the end, if you're a bad citizen of the internet, you are going to get blocked."
- 16:58
It's an inevitability. You can see this on LinkedIn. If, if you have an agent that's using LinkedIn, LinkedIn measures how many actions you take per minute. And if you're violating that, you know, you're going to get stopped.
- 17:07
So we advise our customers is like, "You need to be a good citizen of the internet first. You need to try and obey robots.txt. You need to be careful what you're doing.
- 17:14
And if not, like you're going to have a really hard time." And, and no matter what we do at Browserbase, we can't stop that. We can help with the simple things, but if you're doing something that's against the law, unethical, we don't really want that on our platform. [upbeat music]