WebMCP is here (and you should care)

And I think, Wes, you make a really good point about the multimodality of it. Not everybody's gonna be using the web the same way. Like, we we've known this forever. I think some people are like, the web will go away. It'll just be people using it in a headless way, and then people go, no. Nothing's gonna change. And I think the truth is, like, these experiences are gonna be done all sorts of ways. Like, even a shopping, I'm shopping as, like, a leisurely activity, then I'm gonna be a human in the loop in that activity because that's part of the point. But if I'm just trying to get some shirts for my son that he knows the size of and he knows what he wants, then that I you know, I'm okay with an agent doing that for me.

Welcome to Syntax. Today, we're talking about WebMCP. We got two folks from Google here to talk about WebMCP, and I personally am really excited about this. It's starting to make its way into clients, and I don't think a lot of people really understand what it is or or as a developer, how you should be integrating this into your site if this becomes a thing that everybody implements. So I'm stoked to talk about it. With me, we have Dominic Faralino and Sarah Drazner. They are both from Google working on Chrome. Most people listening to this, I would say everybody listening to these episodes knows what an MCP and MCP server JS. Right? Everyone's slapping them into their agents.

Can you just give us a rundown of, like, what is WebMCP, and and how is that different? Yeah. I mean, I think people get confused a little bit because of the name.

WebMCP is on the client, not the Vercel.

And a lot of the kind of similarities are when we are thinking about the principles, like exposing tools to agents, things like that. But they're actually pretty dissimilar. The governing bodies are dissimilar. One is the Agentic AI Foundation in Linux. That's the, MCP. For WebMCP, it's w three c, like other standards bodies.

For WebMCP, you are really only working on it on the client, but that doesn't mean that it can't, negotiate or create server actions, like, kickoff arrest or GraphQL API interaction.

And then also the way that people are using it agentically.

You do have headless scenarios that you can use WebMCP, but we're also thinking about a new surface area of cobrows. And cobrows is where we're spending a lot of time, which is when you have, like, a human and agent on the loop collaborating on a web experience together.

Thomas, yeah. Yeah. No. I think that Wes that that was perfect. Like, I I think the way I try to describe it to people is, you know, when when you when you give, like, act like MCP tools, traditional MCP tools to your agent, it allows you to just, like, enable the agent to do, better capabilities inside the chat, but, like, we don't have a really like, an MCP for UI, like, for, like, using the site. Like, I'm on Google Sheets. I wanna do this complicated thing in the UI. I can't remember where all the 10 clicks that the functionality is hidden behind. Sorry. My cat's in the background here. And, and so, like, I wanna be able to talk to my agent and say, like, okay. I really wanna I wanna, like, add conditional formatting to this column, and, like, can you can you do that for me? Like, if we've ever seen agents interact with websites, like, Dom and and UI right now, it's, like, really, really slow. They're, like, screenshotting. It sounds it's awful. Yeah. It's such a terrible experience. It's expensive, actually, too. Right? It reenters the inference loop every time you gotta click a button and do anything.

But all the the the agent is thinking in terms of sort of atomic tasks. Like, I want I want to add a video track to this YouTube, you know, editor, or I want to do something in this column in Google Sheets. And all that code actually exists in the ESLint, and WebMCP is sort of, like, finally a way to let the agent get access to it and actually get get to that code and and, you know, translate its its atomic tasks that it's thinking of into actual tasks that the the platform or side that the the client app is capable of handling. Yeah. It's like like, if you look at, like, where we're at right now. Right? Like like you said, if you have an MCP server for something,

and you wanna interact with it, you're you're you're typically just getting, like, chat back and forth. And and a lot of cases, that that's awful. So Node we have, like like, MCP UI where, like, you can, like, have these little widgets that make their way into the chat. And and that's okay in in some cases, but in some cases, like, you know what the best UI is? The existing app is the best UI for that. You know? So, so in that case or if there's not, then you're like, okay. Well, like, hey, agent.

Use this browser. Or, like, the the site owner will strap some crappy little chat Bos onto the side. I don't want that either.

You know? And and, like, I've been I've been, like, beating my drum saying, like, this is, like, clicks and clankers. Meaning that, like, I just wanna use the UI myself like I'm normally doing, but then I also wanna just be able to do agentic stuff with that with that application, with that website, with my own same functionality. Yeah. With the same functionality. I wanna see it. I wanna say, like, another example is, like, if I have, like, six expenses in my bookkeeping software, I don't wanna click them all and edit them all and and select the taxes for every single one. In in some cases, I do wanna click things and drag them into wherever they want. But if it's, like, a batch job, I just wanna say, like, find all the expenses from my cell phone carrier and mark them as having Canadian taxes, and then it will just go off and do that for me. And then I can visually see what is happening in the UI. Yeah. Well and I think, like, one thing that people don't often compare it to, which they should, is DOM actuation.

Like, if you don't have something like WebMCP tools exposed, the agent is inferring what it should be doing from the DOM, from the accessibility tree, from screenshots

and things. This actually gives tools to developers to be able to say, hey. This is what I want an agent to be doing with my site instead of it just happening to them. It definitely lets the developer kind of, express the the capabilities that it knows an agent is gonna wanna get at from its app directly to it so that it doesn't have to constantly do this, like, inference thrashing and the DOM actuation and the accessibility tree reading for every repetitive action and check if the button moved and all it did and then what happens. It can kind of let it let it think in atomic actions and also act in them too. And if you want to see all of the errors in your application,

you'll want to check out Sanity at sentry.io/syntax.

You can sign up today and get two months for free. Century is just a really incredible tool for not only tracking your performance, making sure application has no bugs, but even just seeing what goes wrong when something goes wrong because things go wrong all the time when we're coding. And you don't want a production application out there that, well, you have no visibility into in case something is blowing up, and you might not even know it. So head on to reduce entry dot I o forward slash syntax. Again, we've been using this tool for a long time, and it totally rules. Alright.

Yeah. That's such an important thing. Node thing I when I talk to Node folks about WebMCP, and it arriving in Chrome, people say, is this a Chrome thing? Is this another Google thing? You had mentioned that there it's involved with the standard support. So can you touch on,

the status of all that, and is this just a Google thing? No. It's it's definitely not a Google thing. We we're working with with other folks in the community, other model vendors. We're working with Microsoft on this.

We're developing the API as an open web standard in the w three c. It's managed in the what we what is currently the Wes machine learning community group, where some other agentic kind of APIs are for the web platform. And we've been in close contact with model providers, and also developers who are gonna have to use the API and other browsers that we'd love to have implement the API. So there's been lots of discussions in in the normal standards, bodies, gauntlet of consensus and and controversy, with with other vendors and developers to to get that push forward.

And so I think that's, we're we're still early stages in all of that all of that sort of thing. But, so far, it's going really well, and we've gotten a lot of good developer feedback. And, clearly, I think the community has has proven the the the need for something like this. Like, every time we present this to, you know, a model vendor or, like, an extension provider that, you know, augments Chrome with some agentic capabilities, they're always like, this is what we what we want. Like, we want an actual, like, clear capability layer to talk to the site instead of have to, like, burn the user's wallet through inference loops and screenshots and all this stuff that takes way too long. Yeah. And I know you guys work at Google, but, like, the the idea was is that, like, whatever

client you're using. Right? Like, if you're using like, it's it's in, ChatGPT desktop right now. If you use the built in ChatGPT desktop browser or I think the the Chrome extension or whatever, it will surface those tools via, like, we'll talk about the API in just a second of how you JS developer can implement it. But it's it's any client. Right? The the hopefully, eventually, will work in, like, the cloud desktop or Gemini or, like, any client that wants to be able to to do this. It should be able to just open up a browser and say, ah, here are the tools that I have available to me. Yeah. If people haven't seen already, ChatTP

ChatGTP and OpenAI came out with a big announcement that they're supporting WebMCP as a first class citizen. There's a big hackathon.

We're involved in that even though, you know, it's across many different companies like Vercel and, CloudFlare and a few others. So, yeah, it's more an ecosystem thing. Edge is definitely proponents of the spec. They're worked on the first versions of it and things like that.

And I think the Chit Chat GPT sort of hackathon that they've announced with WebMCP is gonna be a really good way along with the origin trial that we're running in Chrome. It's it's a really good way for us to pressure test the API and understand, like, how developers are actually using it, what granularity of tools are they writing. And then it also puts a lot of, importance on the ability for the ecosystem to adapt and and come up with, like, real evaluation metrics and determine, like, I'm I'm I made this tool. Did it help the agentic path on my site, or or did it hurt it? And, like, I made this change to my tool. Did that help, or did that hurt? And so we're thinking in terms of, like, ecosystem activation to understand what it takes to write a good tool and what the API needs to do to expose tools, you know, in in the most effective way to the to the agent brain so that it can it can have a really effective experience on the site. And so we've seen a lot of feedback from our origin trial and and, this hackathon coming up, that's gonna help inform the API shape and all of that stuff. So the API is

and and the last time I implemented it was both an HTML API and a, like, a declarative and an imperative API and a, like, a JavaScript API.

So how somebody has a website. They are a JavaScript developer.

How do they implement this into their site so that agents can use their way their website? Yeah. I mean, I try to think about it in a couple different ways, but this is also why system design is really important. You want kind of task level milestones. I think a failure mode can here can be that people kind of attach to every little micro interaction on a DOM element. That's not exactly what's intended here. We're thinking about more, like, stateful or or potentially effectful kind of tasks that an agent can use.

We're actually collaborating with some large frameworks to investigate how we might infer tools just from your existing application authoring. So you could just write g s x, t s x, and then there's a separate potentially, like, reconciler pass that infers some of those tools that you can look at.

But if you're working with frameworks like most of us are, you wanna expose tools potentially via hook. I made a a WebMCP tool hook, that's on Npm that us in Chrome will keep up to date with the API.

So don't leverage the existing DOM structure.

Think in agent actions.

I would also wanna leverage existing APIs in your application, like REST or GraphQL.

If you have those, you can those can be invoked with a tool with a really good description.

We have a lot of, annotation hints, and those are also really important.

So, one really important part of the spec for me is that you the ability to keep your site secure. So, you kind of also want to go through and look at, are you exposing user generated comments somewhere on your site? If so, you'd want a read only hint. You'd want, user generated content ESLint, because the agent can then say, okay, I shouldn't look at any of those and take action. That can guard against things like prompt injection attacks, or, you know, agent traps, or things like that. So you might wanna think through, if you're a site developer, like, where are you taking content that you're not sure somebody might be doing some of those harmful actions in?

Yeah. The security side of it is is is pretty, pretty vast. I mean, like, you know, historically, when we work on web web platform APIs, we kinda have, like, two consumers. There's the browser, and then there's the developer, which is, you know, writing experiences for the user.

And WebMCP is unique in that, like, whenever we adjust the API or add new things, we have we have to make sure it works well for developers. Like, does it make sense for you to add this attribute on your tool, or, like, is it okay if the tools are unregistered at this point? But then we also kinda have to ask the inverse questions to to models Wes, you know, it's like, can you does your security Yarn, like, benefit from consuming this attribute? Does does having a tool Yarn itself as read only or untrusted? Is that useful for for a security harness and an agent? And, you know, if a tool unregisters here, is that too much dashing, or do you guys have a caching mechanism? And so there's kind of like two different halves of this API, and and getting feedback from both sides, has been has been super important.

And like like Wes mentioned, there's the imperative API, which is just ordinary JavaScript, allows you to basically pnpm in a a JavaScript callback as the implementation of a tool so that the model can sort of call directly into JavaScript that the developer chooses to expose. And then there's the declarative API, which is like annotations, like like attributes on on existing forms.

It's probably no surprise we find developers most interested in the imperative API so far Wes most of the site code lives. But I think declarative also, may have an important place because it it's it's an easy lift for your, like, you know, crappy local, like, government car registration website that is, like, from 1995 and has a bunch of forms on it to just, like, sprinkle a couple things on and then kinda become a little more agentified and Yeah. And allow, like, those kinds of use cases. So, yeah, we're trying to collect feedback from developers on both API shapes and see what sticks. Yeah. And so many of us thinking in interfaces

anyways. Right? These are the interfaces that the users are interacting with. And if we view an agent as another user, you know, just being able to sprinkle that declaratively on top does feel like a big win.

Yeah. The I see, like, there being just, like, a 10 stack query plug in for this type of thing. You know? Because, like, I already write all my all my mutations, all my queries of, like, get items, update item, delete item, you know, all that all that crud that that we're we're used to. And then I I I in most cases, I I probably want, like, 90% of those also to be available to the agent so they can just programmatically call it. So, like, I I see a a place where

you maybe aren't doing a whole lot other than just throwing in, like, a plug in or a hook or something like that. Yeah. And, I mean, it also allows us to you know, if you do exactly what Wes is talking about, you can also return structured errors back to the LMs. You know, you plug in exactly the what success happens and the reason, and so that the agent knows what failed. So you already have existing, you know, mutations, actions, dispatch, kind of logic, you can leverage that existing logic for some of your tools.

What, like, what kinds of apps are are like, you see, like, the most obvious benefit from this type of thing? I know it probably can be used for for anything, but is there any, like, specific use cases where, yeah, that makes a lot of sense. Well, we're collaborating with people like Shopify has integrated.

So we're seeing a lot of use cases for Shopify. You Node? Obviously, ecommerce is a big one.

Right now, we're talking with YouTube about their in page agent being able to surface, like, if you ask YouTube, all of the things that people might not normally get exposed to that are on the YouTube page, like playback settings, and, I mean, that's one that people usually find, but there's all sorts of settings in YouTube that people don't necessarily find easy Sanity, so anything where discovery could be limited, and it an agent can kind of surface those discovery mechanisms.

Instacart is another one where people probably want to, like, identify or make their shopping experiences a little bit easier. Things like that, we're seeing a lot of, benefit from. So yeah. Yeah. And I'm sure you're gonna uncover a lot too with this WebMCP

challenge that's going on.

So we have this new WebMCP challenge. The deadline is September 3. So by the time you're hearing this, the deadline will have passed. But it seems like there's a massive amount of submissions or at least people signing up to be potentially submitting this already. I I would imagine you're gonna see some really cool stuff, and Sarah, you're a judge for that. Right? That's right. Yeah. And we're looking at, like, creative uses for, WebMCP.

We're looking at for the good of humanity, we're looking at, do you structure your tools well and make sure that everything is, you know, kinda copacetic and easily surfaced by an agent? So there's a number of factors we're taking into consideration with the judging. I am really excited so far. I think last I checked, it was, like, 5,000 entries that I've already gone through. So Holy smokes.

Yes. One of those will be mine. So, hopefully, my tools are are set up correctly. I'm I'm expecting to submit my Wes of the the demos that you gave Wes, like, a three d modeling software, which three d modeling software and video editing software, I think, are two huge use cases for this because those are UIs where, like, I'm I'm not using a chat for that. No chance. But, like, we all of us on Syntax, we use DaVinci Resolve, and we all have the DaVinci Resolve MCP running because in some cases, it's way easier to just type in the box what you want it to do than to figure out what the crazy clicks and and whatever that you need to do. And I love I I've been calling it you you're calling it co browsing.

I'm here to tell you it should be called clicks and clankers.

I mean, that you you know, because we had we had bricks and clicks when the, like, the web was coming. You know? You have the store, but you also have the online thing. Now we have the human and the agent. Clicks and clankers, nobody seems to be it's not catching on, so I'm reaching out to you

to for that help. We'll make it happen. But, no, I I agree with that that general use case. Like, I I do a lot of, like, makerspace stuff and and, like, you know, CAD modeling. And, like, there's there's only so many times I'm gonna, like, relearn how to do the same thing in, like, Fusion three sixty. And, like, I'd really just want the chat to be able to, like okay. I can describe exactly what I want. I can kinda probably get there on my own, but it'll be quicker if you could just, you know, take my intent and map it to actions on the actual site. And so I think that kind of use case where these complicated configuration UIs but I also still need to be involved because I need to, like, double check the end result or see the measurements or or whatever.

That that's I think the most the most, like, immediately useful. But but like we said, it's also pretty useful for headless scenarios. Like, a lot of times if you're just messaging your your, you know, Node bot on Telegram and telling it to do something, it's gonna open up a headless Chrome in the back and talk to it through CDP.

If every site it it uses has WebMCP tools or it can talk to the service worker tools inside WebMCP and do some background executions, it's gonna make even all of the headless scenarios, like, just as easily if if they're if it's not talking directly to the server. So I really I I think it's a unlock for a lot of different use cases, click sync, anchors, and headless.

That headless one because I was talking to some people earlier on in WebMCP, and there wasn't, like, any headless thing Node. But now you're saying there is. So, like, you're saying that I could technically just have, like, a, like, a box running somewhere with, like or or my just leave my laptop open. And if I have my websites open or they can open them, then I'll be able to access those sites that have WebMCP?

I think so. Yeah. Like, if if you're if you're messaging your, you know, your bot on on some VM somewhere and and it's really doesn't present a lot of UI to you unless it really needs your intervention or something like that. And I go tell it to Yeah. You know, register my car or update my my license registration or something like that. And it navigates to the the site. You know, the state of Massachusetts does not have, like, an MCP server. I'm sorry. And so but maybe they we maybe we convince them to drop a couple attributes on their forms on their site. And and now, you know, my bot, wherever it's, you know, living, headless or not, you know, behind a Telegram VM or not, if it can if it can actuate the site through WebMCP tools exposed through CDP, it feels headless to me, and it doesn't really make a difference from the agent's perspective because it's just interacting with the site in in any in any way. So I think that's kind of the idea.

I mean, if you never have to go to a site or whatever to register your license or, like, I would love that.

Yeah. Exactly. Sarah promised me a promotion if we can get all that. I'll put these hands to I'm sure she said to do that. So

This is a perfect use case for something I'm working on because right now, every single time it does pops open dev tools MCP specifically.

I have a a Mac mini right here behind me that you might be able to see it flashing occasionally.

That's because it's running dev tools MCP to export a video from this application, and this JS, like, such a perfect that if I I would never need to see the interface for this. The the process should just be able to do it headless. So,

that's really exciting to hear. Yeah. I mean, the dev tools for agent stuff is really exciting. If you're working with WebMCP, like, sometimes people we get, you know, a lot of people asking us, like, how should I be debugging this? I think DevTools, Chrome DevTools for Agents or DevTools MCP is a really great way to, like, send an agent off to do a bunch of things. It also invokes Lighthouse for agents. It also, can perform audits for you. And so if you're not using that already, it can integrate with a lot of different models, including the Frontier models.

So that's a really good debugging journey. I think some people, maybe not everybody, knows that there's also a Chrome extension that you can use. So I also tend to use the Chrome extension with, if you look pnpm up in dev tools in the application, tab, you can see all of your tools listed. And then you can invoke the tools on the page and automatically see right in the page, feedback for how those tools are getting executed and some observability.

So, like, there's a number of different ways you can do it. Some are, like, ESLint. Some are, like, directly in page, and those are cool. Mhmm. And what what's that Chrome extension called? Is that just the Chrome Dev Tools MCP extension that's been rolled into that, or is that a sep is that the separate Web MCP extension?

Yeah. There's one that's right directly in Chrome Dev Tools. That's the one that I was mentioning in the application tab, and then the one that, Scott is talking about is a more headless model that you can just Yep. Run-in the background and, you know, it kind of operates sort of like Playwright and Puppeteer, which are also good debugging tools as well. Wes, that's awesome.

Another question I had, I don't know, maybe six months ago when Wes first started talking about this was, like like, what about multi tab? And, like, I assume that you have two tabs open, and both of them expose WebMCP tools. Your agent would be able to to use both of those. Right? Like, one example I had is I built a shopping list application where you could add stores, and you can add items to each of those stores, and you can mark them off, yada yada. Right? And then I had another, recipe website open, and and I wanted to get all of the ingredients from that and put them into my shopping. Right? So that's two totally different websites, two totally one Wes WebMCP, one was simply just scraping. But, like, I should be able to do that. Right?

I think yeah. Like, the, you know, the agent will be able to use WebMCP tools where they exist and and make use of that site functionality when it can.

But, ultimately, yeah, it's the agent's sort of prerogative to understand what what origins to reach out to, what tabs make sense to interact with to to fulfill kind of the user journey. And this is actually one thing we've been we've seen a lot of confusion about with WebMCP.

When when folks are reviewing it from, like, a a traditional web platform perspective, it looks kind of like a wonky API because it's it's sort of, like, the first of its kind, like, really facilitating agentic use of traditional web content. And so I I think a lot of people mistake it for like, oh, this is the agents on the web API.

And I think from our perspective, it's it's really like, no. This is the like, let's give developers a chance at presenting something that's sensible for agents because agents are already on the web API. Like, it's agents are using the web, you know, regardless of whether whether WebMCP exists or or not. And so Mhmm. So we've been, like, really trying to understand, you know, like, there's been a lot of confusion, for example, for security.

It it you know, when people think through what it means for an agent to use a WebMCP tool, kind of like, oh, like, an agent can use this tool, but what if it has, like, some some stuff from another origin, like, in its brain, and it wants to, like, share that information with this tool? Isn't and that's kinda, like, violating the same origin policy. Right? Like, that's that's kinda scary.

And that's that's violating Cors. Like, what does that what does that mean? And I I think it's a little hard to think about because we don't mind users violating the same origin policy. I'm I'm the user. I can see all my cookies. I can see all my tabs. I can see all my origin Node.

But that's because the product is is kinda serving me. And so in one sense, the agent is kind of an extension of the user and sort of punches through the traditional, like, web sandbox security model. At the same time, like, users are not comfortable just, like, probably giving the agent its entire identity and letting it assume it's it's full persona, like, on the on the web and just browse and do whatever it wants. And so there's only so much of this we can control from the platform perspective, which has, like, the same origin policy and cores and all that. And we're starting to think about what it might make what what it means to produce, like, an agentic platform, like, in in the product.

And instead of thinking about, like, the same origin policy, what is what is a safe origin policy look like? Some some agentic model and browser so that the agent can know, like, yeah, I should I should be able to assume the user's identity on on these four sites related to this task, but I I shouldn't be able to do everything. I can't read all their cookies. I can't go to their bank and and start making transactions. And so we're starting to think through what what a what a capable, agentic web harness looks like in a browser that integrates somewhat with the web platform and somewhat with the product, to actually make agents on the web secure and also facilitate the use of WebMCP

through traditional web platform content. So it's a it's a complicated model. But Yeah. And, like, to zoom out for a second. So there's you know, as Dom said, there's only so much you could do on the platform side and for site developers, but we're also talking to agent developers and our own agents about what's potential there. And I think one thing to get people to really understand is that we this won't be solved by one thing. It has to be a multi layered defense strategy because agents can be somewhat Node deterministic, and also you need both the site side and the agent side. So some things that we're thinking we're doing within Google that we're thinking about open sourcing and making more of a standard are things like prompt injection classifiers, so they can identify attacker instructions in content before instructions kind of go out, critique LLMs. A lot of people know about, like, a secondary judging LLM. You can imagine that being applied to, like, web surface areas.

And then also just, like, restricting origins. Like, you may wanna have, like, that LLM that's a judge say, okay. You were supposed to go book travel for me, and you can go to Expedia and United, but why are you going to my bank? Why are you going to my health site? Like, to make sure that we're not going off to Origins that they shouldn't. And finally, there's, like, this kind of special agent containment layer that we're thinking about, exposing to the community so that agents can have, like, more of an identity that's separate from the user. Like, so far, the agent is you, but you could imagine that you also you might wanna have an agent identity that's separate from you that, like, can only spend $20 a week,

only has access to some information about you, only has some passwords, things like that. Here's $20. Yeah. I gave my Shopify $20. They the Shopify has, like Go get your Shopify. Shopify website, they have, like, a /agents.md with all the information about how how to to communicate with it. And I you have to, like, explicitly give it a little bit of money and let it rip. I thought that was funny.

Nice. That's great. Oh my god.

Can we talk about something that makes me sad about the web is the performance where we went in, like, the process of, like, a year and a half. We went from it really matters how quickly your key ups happen, and you should not block the thread, and, that we have all of these web vitals about making everything super fast. We cared so much about all of that stuff. And then these agents came around, and it takes, like, $4 and three minutes to click a link to do something. And I was like like, this, the experience sucks. If you're if you're if you're out somewhere and you don't really care, but when you're waiting on it to do work, that experience is absolutely awful. Is there is that obviously, WebMTC will get that better, but will we bring vitals to agents?

Yeah. I am actually kicking off a, like, what would core web Vercel look like for an agentic web thing.

So there's things like, you know, if you're using, you know, Claude or Gemini or ChatGDP, that time to first token is a new metric that everybody's kind of looking at. And that kind of those kind of streaming deltas, because you don't have just like the second it goes, but you also have the second that it, like, streams all of the possible input.

But in co browse or wait. What was it? Clicks and clankers? Clicks and clankers. Clicks and clankers. Okay. I got it. In clicks and clankers approaches, you the the thing that's fascinating about it is that some of the old Nielsen things still are, like, are still some things that we're seeing here. So, like, one second per tool call, five seconds before you wanna see an entire action go through. If you're not familiar with that, those are some of the earliest human computer interaction things. Like, people don't wait longer than five seconds for a Wes page. That actually has worked in, like, clicks and clankers approaches JS well. But the thing that we didn't have in there is that when you're watching co browse, you're not just judging the agent and the time, you're seeing how fast it is compared to you. Mhmm. Because you don't want to be doing an experience on the web where you're, like, watching it, and you're like, I could be clicking that faster. I could be doing this whole flow faster. And so that's the first time we're seeing metrics that might be comparative to a human.

And the real trick is exposing that to developers. Like, in the past, like, we we could let developers do things on their site, and we have, like, performance observer and that kind of that kind of thing that lets lets the developer get access to how long that navigation took and how long that animation transition took. And they can understand when they make changes on their site if it affects the user experience through real user metrics.

But the agentic side of things is a little trickier because, like, a lot of times, the the success metrics are kind of locked up in the agent's brain. It knows, like, how many model turns it took to fulfill a user a user journey and how many user journeys, you know, it it you know, warp completed with a WebMTP tool and how many tools were involved and how many tools confused the agent. And so we're we're also trying to think through we're super early in this, but we're trying to think through some ways to expose some of those metrics to developers as well, whether it looks like some agent performance observer or some reporter API or something like that, where we can let developers actually measure the effects of the AgenTic targeted things that they're doing on their site so they can know if they're actually having a a positive or negative impact. Because right now, like, it's it's slow for a lot of reasons, and WebMTCP tool calls make that a lot faster, but some of them are still slow. But one of the, like, really big challenges here is, like, the whole thing's opaque. Like, we kinda really have a hard time measuring, like, how how would changes I make to a tool impact the success rate and how they impact the latency and what's confusing the model or not. We need the developer to

transparently understand that kind of stuff on their site, or else they're flying ESLint, and it's gonna be impossible to make changes and and measure against them. Do you also have a way to, like, measure just for, like, regular people using it as well of, like like, success rates and whatnots? Like, I think back, I was booking a hotel on Expedia a couple weeks ago, and, like, I wanted to buy I want to book, like, a suite that a separate room, and that wasn't, like, a filter on Expedia.

And I was trying to do it entirely agentically, and I was like, this is awful. I need a map. I need photos of it. I need, like, all of I need the UI. It it's not a very good experience. Right? And, like, all these tech bros are just like, oh, yeah. I booked a flight for me, and I bought red shoes online. And, like, that that's not how regular people do do their work. Right? So, like, is there some sort of,

like, measurement that you're doing with, like, regular people as to, like, whether this is something they use and is sticky enough? I do think that for some of the, like, crux things that we're working on for agents, we are invoking like, we have a bunch of UX researchers who are looking into this so that we do these types of studies both with, like, real users and then also, by doing analysis of the Wes. And, like, we have a lot of data because of Chromium being used by so many people that we can kind of leverage here. I would say that we're pretty early in, like, full Scott, this is the way that everything works. We do have some targets. And like Dom said, I don't think that anybody's at those targets yet because these experiences are so new. People aren't even used to building out a product experience that might incorporate something like this flow. Like, that even just, like, for PMs of a site to, like, think through, okay, what does that look like if somebody is going through a cobras flow? Sometimes we've had these, like, deeper partnerships, but we do need, like, site developers and product managers and everything to incorporate that kind of thinking as these, you know, new agentic surfaces evolve. And I think, Wes, you make a really good point about the multimodality of it. Not everybody's gonna be using the web the same way. Like, we we've known this forever. I think some people are like, oh, the web will go away. It'll just be people using it in a headless way, and then people go, no, nothing's gonna change. And I think the truth is, like, these experiences are gonna be done all sorts of ways. Like, even a shopping experience, if I'm shopping to, like, you know, JS, like, a leisurely activity, then I'm gonna be a human in the loop in that activity because that's part of the point. But if I'm just trying to get some shirts for my son that he knows the size of and he knows what he wants, then that I you know, I'm okay with an agent doing that for me. And I think a lot of these things are gonna evolve as Wes you know, XR, whatever experiences

involved with it. Like, oh, computer, show me what I look like in this shirt. Okay. Now this shirt. You know? I mean, like, who knows Wes all that stuff. It's gonna evolve as we go. Yeah. Totally. That's definitely one of the challenges of API design and and this kind of initial era. Like, all of this agentic web space is is super nascent,

and so it's it's it's hard to tease out some of the patterns, you know, we're seeing and and, you know, derive what experiences we can from from actual MCP and see what makes sense over in WebMCP. And, you know, everything's everything's new and and moving so fast with with from the developer side and from the model side.

So it's kinda hard to pin this down and understand exactly what what makes sense, which is why the origin trial and the hackathon and that kind of stuff is is a really good way for us to collect experience.

Have you heard from any site owners who are, like, resistant to this type of thing? Because, like, that's, like, another weird spot is, like, if I'm an airline, I don't necessarily know and this is more like the MCP server way. If I'm an airline if you ever tried book a flight, they try to hard upsell you on absolutely everything. They well, what if you get sick? You $20 for for the insurance and all of, like, these, like like, black tactics.

Dark tactics. About to be defensive about that. Right? Oh, we thought you might need this, so, therefore, we added this. Yeah. So, like, I'm I'm wondering, like, the airlines doesn't just wanna be like a like a straight up utility

for just vending out the cheapest flight for this type of thing because they wanna be able to make make more money as well. So have you heard any, like, pushback from people who own sites and are like, we don't want this.

Yeah. There are a couple of cases that I probably can't disclose on a podcast.

But, I do think that there's, you know, when I look at the interest, it's far more people wanting to expose WebMCP and tools because they want Mhmm. The agents to be able to discover and not fail. And, like, I think what they're really worried about JS, like, oh, okay. If it's just dom actuation and things like that, then we can't guarantee that they're gonna have a good experience.

They might go to some other place or something like that. So, mainly, it's interest. But there have been a couple of outliers of people wanting to be like, maybe we just say, everybody go away, and, like, abuse in that kind of direction.

I think in cases like that, we are still, like Dom said, we are so early on in trying to figure out what those loops and experiences might be. Do you empower the user? Do you listen to the site owner? Like, if the user really wants to be using an agent, are you going to flat out tell them no? Like, I don't think that the industry has a collective answer for that yet.

But typically Wes try to, like, put the user first.

And so that, you know, is a little bit of attention. Thankfully, it's not that common.

That's good. And of those ones that are common, I I wonder how many of them are just trying to protect their business. Like like, like, either, like, come along for the ride on the agent world or or, like, be left behind, a lot of people are saying. So I I don't think that's all of them. I think there's certainly a lot of people who can, like, like, I guarantee Amazon could say, nope.

None of this, and and that would be a big problem. Same with, like, Apple Pay at Walmart. Walmart just says Oh, yeah. Nope. You know? And, like, that's a big deal. They they're big enough to do that. But for a lot of people, they're they're not big enough to actually to push people around like that. I did not know that Walmart does not take Apple Pay. That's that's news to me. They have in Canada for years, but Node is a is a great country. I think they just have a lot of US is just getting it now. Yeah. Just getting it. Yeah. End of twenty twenty six. Wes.

It's a hard space, though. Right? Like, should my agent be, like, watching ads for me if that's, like, what the site wants? Like, you know, probably probably not. Or, like, should it be clicking on ads that it thinks I'm useful? Like, there's a whole monetization model that's, like, totally nascent here about, like, what what does agents on the web even look look like for the traditional funding and attribution and refer and all that all that kind of stuff model. Like, that is that's like a whole new space. It's it's beyond WebMCP. I mean, it's WebMCP play plays a role in it. But I think, you know, we see we see a lot of, different corners of the industry rallying to answer similar questions. Like, this is kind of, you know, also related to the UCP, the the commerce, protocol spec. You know? Like, there's there's a lot of upselling there. Like, how how does that integrate? You know? There's gotta be answers with that.

I don't know. Maybe maybe someday, there'll be, like, an ad viewing spec where you can view your ads through m speed. No. Just kidding. But, like, you know, it's there's a lot of a lot of new things here. And then and then, obviously, the the natural thing is for, like, large kind of, like, business conservative enterprises to to maybe resist it and and keep their traditional model. But I think, ultimately, most most, vendors of that sort will end up figuring out a way to integrate industry solutions to to, you know, enable new ways of monetization and new ways for agents ESLint interact with their sites. But exactly how is is unclear yet. But, yeah, it's it's gonna be an interesting future for sure.

We are working on some things internally, but I don't think that they're totally ready for prime time. Maybe we, you know, send you all links in the future or something. Alright.

Send us our way. I didn't realize this. So, like, for people listening, that's ucp.dev, Vercel commerce protocol.

And then I've also been keeping my eyes on the x four zero two project, which is the, like, agentic payments. I know CloudFlare is rolling out wallets soon. I know Stripe JS has their wallets in The US, which the your agent can spend money on. So it's that's obviously not part of WebMCP, but it is kind of related as to how do agents use the web without bankrupting everybody.

I mean, we do examine this as part of health of the web because in order for people to keep the web healthy, they need a way to make money off of it. So if you have, like, Beyond Commerce, if you have a content site, that makes money off of ads in order to show and display content, then you're kind of going warp more towards a sub why can't I say subscription? A subscription model for those things.

Yeah. Totally. Yeah. I mean, I mean, because the AI agents are just slurping all that stuff up now, and there goes your your income. So yeah. Yeah. It's a

different world, I think, for a lot of sites that are are trying to work on that model right now. Yeah. And I think it's definitely I think it's important that the web evolves to try and meet meet the moment. Like, I think there is a lot of, you know, understandable and and natural pushback in general about the web evolving too fast or adding too many AIs that are tailored to agentic experiences. And, you know, these things have a necessary long tail of controversy.

We have, like, the WebMCP and the prompt API and and and and so forth.

But, like, I I think it's ultimately a good thing that that we're focused on trying to figure out what the web's real place is in this kind of agentic world. Because, like, I'd certainly rather us all argue about a healthy web that is still relevant than a dead one that died because we we didn't keep up with any technologies that are that are coming down the pipeline. I think that's, like, a really important thing to worry about if the web doesn't really meet meet the AI moment and and figure out what it means to present kind of kind of two platforms now. Right? Like, traditionally, it's it's always presented the web platform to developers and to the user. This is things like readable streams and module scripts and anchor positioning. But, like, now Wes probably need some, like, agent platform side of sort of things where, you know, I I kind of envision a future where you can plug any agent brain into your browser, and then that agent brain can sort of get, like, whatever containerized access of you know, to your to your your site data or things that are useful to you as the user and be constrained by the the guardrails that the the browser's agent harness kind of is able to provide.

And then whether that is partition credentials or read only views of of certain, you know, parts of the user's personas that they can actuate on their behalf and so on. Like, all of these kinds of things, like, Yarn gonna have to you know, we're we're currently working on them, and we're currently trying to figure out what they're gonna look like in the future and understand what it what it means for the agent to kind of have a or Scott of the browser to play a role in sort of, like, this agent platform, like, space. Because that's kind that's very new. Like, we've only had platform and and product.

And now, like, product is is kind of containing, like, a bunch of agent primitives. And we're actually thinking about this from the extensions point of view on on Chrome as well. Like, it's it's really widely known that, like, a lot of agents that are living in extensions on on Chrome, like, they they need, like, full access to the page, and so they they get, like, accessibility tree and screenshots and all this kind of stuff. But to do that, they kind of end up tripping over the debugger API, and then it kinda flashes this, like, scary banner, and enterprise clients don't don't like it because the the agent has so much direct control over the page through an API that was never really designed for agent usage.

And so we're kind of stepping back and being like, what does it mean to, like, refactor and redesign all of these things to provide, like, an actual agent platform for agents to plug and play straight into your browser, but but meet the safety standard that the browser is known for through whatever safe origin policy or harness or tool tool set or read only view of the user's state or data that that makes sense. And I I think that that's, like, one really big thing I think is important for browser vendors to to focus on. And it's actually it's the reason why we see a lot of these new agentic browsers sort of come and go from other companies. Like, they're, you know, spinning up their own their own binaries and trying to encourage users to to use them as their as their browser, not because they're they differentiate themselves on the on the web platform's perspective, but it's because they bring with them a bunch of agent tools that traditional browsers might not have thought about, you know, from from day one. And so I think it's it's important for all browsers to understand, like, what it means to kinda be an agent platform and and a harness. And I imagine, like, a a large suite of plug and play configurable tools and and security policies Yeah. In any kind of browser that manages an agent.

I mean, it's a Dom's point. If you don't think about it, then you can't secure it, then you can't make it private. Like, whether or not you want agents to be on the web or any individual browser wants agents to be on the Wes, they are on the web.

What security has been talking about is this fourth actor. Right? Like, you have the platform, the user, and the site. And then all of a sudden, you have the fourth actor, which is AI agents. And they call it, you Node, in security we call it a trust diamond.

And that you can't think through those pathways and actually make them safe and secure unless you're actually paying attention, like the you're not hear me. I can't see you. It's not, necessarily the approach that's gonna allow us to make things safe for users.

So WebMCP supported by OpenAI, Chrome, Shopify, Vercel, CloudFlare, have you heard any peeps up or down from the the the two boogie monsters, Anthropic and Safari or or Apple?

Firefox and and and, Safari or or Anthropic and Well, Anthropic

because this would have to be in cloud in order for it to be, like, everybody to use it. Right? They're such a big player at least right now. And then, like, Apple and and and Mozilla as well, I guess. Yes.

Yeah. We we have heard from all of them. Like, we've been talking a lot with with Mozilla about the imperative API. They're they're pretty interested in it, and they they've expressed some some public support to the the imperative side of things, from WebexDB, which is awesome.

We've been in discussion with, yeah, other other model providers like anthropic and browsers like Safari. I think probably we can comment on some of the stuff that's already public.

I think there's a lot of enterprise interest from from both parties, because this is where, like, a lot of knowledge work happens. And so making sure that there's widespread adoption among developers that that can help enable WebSocket in these in these products makes a lot of sense.

And,

yeah, Sarah, do you have anything to add to that? Yeah. So we are talking to all of the people mentioned. They are investigating what it means for them. And so JS part of that investigation, are doing due diligence on what they wanna be supporting, what they don't wanna be supporting, and formulating thoughts. And so I would say we're still in the kind of, like, meeting with people and talking through things stage of things, and not in the, like, here's what we can formally announce stage of things, but it is covered territory.

That's cool. I had used a, like, a WebMCP b, extension, which basically turned my WebMCP websites into, like, a, like, a proper MCP server. And that was really cool because then I could just I could take that MCP server and put it into anything that supported MCP, and and then I could I could just chat with it. Right? That was cool because I I did. I slapped it into Claude. I typed into Claude, but I could see it controlling my browser.

So I thought, like, maybe that will be an experience at at one point as well even if they don't end up supporting it.

Yeah. I just talked to him last week.

He's really a like, I don't know if you've talked to the creator of WebMCPB.

I think he does a really good job of, like, thinking through what people might need because he used to be a consultant for all of these kind of, like, Salesforce and other companies.

So he's kind of good at those, like, glue layers. We are incorporating a polyfill, that, you know, he worked on previously and things like that. So he's a really good community member, and I like the work that he's been doing. This is,

Alex Nejas. Yep. Yeah. He's been great. He's he's been an awesome partner.

He he was one of the original folks, that came up with kind of the original shape and idea of what WebMCP might look like, and he's pretty active in the community group.

It's it's been great working with him and, you know, building building polyfills alongside of him and so on. So yeah. But but you can definitely imagine that any any model vendor that, you know, has a Chrome extension or or really wants to be able to perform knowledge work tasks and user wants, you know, for users, really, really wants something like this. Like, it's it's amazing. I mean, every time we we we bring it up to people and and and they they're just their eyes light up, and they see that all the benchmarks go green, and they're kind of like, oh my gosh. Yes. This is this is saving time, dollars, latency, everything. And there's already been a bunch of public, like, benchmarks about this that actually integrate that kinda build their own harness sort sort of like MCPP that sits in between WebMCP tools and the cloud code harness or the OpenAI extension and so on. So you can actually get a feel for, like, what it would what it would look like for these extensions and model providers that don't support WebMCP today, like, what it would look like for them to actually do so from a performance and a usability perspective. And, so far, it's it's been it's been really good feedback.

Yeah.

Is there anything else that we haven't hit that you all want to make sure we cover?

I think I think the one thing I'd I'd love to to encourage is the community to to keep keep tabs on the w three c spec.

You know, every every day, we get new issues filed, bugs or proposals or additions and so on, straight to the repository. It's a really good way of getting real world developer feedback, and and real world model vendor feedback. Like, we we have extensions, you know, folks that that build kind of like, community, DOM actuating extensions.

They love to chime in on the repository and help us understand what parts of the API make sense for their harness and what doesn't. And so we would love people to just stay stay in touch and keep in the loop of the API and and also provide their feedback to us because it's all super useful. Like like we mentioned a few times, this stuff, it's a really, really early space, and we're we're figuring it all out, and we're trying to make sure what what ways we can impact the security and the usability of the API.

And so I would encourage people to to stay in the loop and start experimenting. There's been some really, really great tools being built.

There's o r a dot a I, which is a tool that has been built by one of the MCP ops, MCPUI cocreators.

And this this kind of, like, helps you measure how agentic journeys are are, like, actually happening on your site and how WebMTC tool calls are being used by real world agents. And it's a good, like, benchmarking, kind of framework that that we've been, you know, looking at and thinking through for to as as a way to understand how useful tools, like, really are. And so, the more the more feedback like that we get from the community, tools being built to measure how good how good, WebMCP tools and agentic experiences are, all of that, like, really helps the industry create something measurable and and effective from the use of Scott just WebMCP, but anything agentic on their sites. So we would encourage developers to get us involved in in that sort of space as possible because I think it's it's super useful feedback for the browser engineers and,

Node vendors as well. That's great. Yeah. Like, seriously, folks, listening to this, try build something. Slap it in your site. Try build, like, a little to do app or whatever and, like, and, like, give your feedback Node. Because, like, I'm one of my first pieces of feedback was, like, I want this to be headless as well. And, like, I don't I don't think that was me, but, like, the the now it is. Right? And, like, provide your feedback now so that we can, like, nail this.

Because even if you look at, like, the the journey of MCP, it's had so many high highs and so many low lows. People have been said it's over, like, six times since it's been released, and it's just because we didn't know what it needed to look like. So, like, chime in, and then that's super helpful. That was so funny. I was I was at the MCP Vercel Summit,

conference in New York City in in March, and, you know, the whole this was like when skills were happening and and CLI Yeah. Kind of popping off, and everything was MCP is dead. And so they actually threw an after party on one of the nights at at some, like, kind of like Yarn, nice cocktail bar in Manhattan. It was called MCP is dead, the funeral. And they had, like, a quartet, they were, like, singing sad songs. And it was like, they went so far. And so I'm like, oh my gosh. But, yeah, yeah.

It's it's entertaining to see see all the community hype and and unhype about random things as as they fluctuate.

But, yeah, otherwise, thanks for giving us the opportunity to to chat about WebMCP and kind of, you know, talk through what what we think the future of all of this stuff might might look like. You know, we're we're marching forward as fast as we can, but, also, a lot of this stuff is is pretty early and and speculative. And so, you know, we'd we'd love to stay in touch and and,

keep keep an idea on what developers are building with with these tools and how they work in public. Amazing. So now is the part of the show where we give you the opportunity to share something that you're really interested. We call them sick picks. These are things that are just in general that you're enjoying in life right now. Could be literally anything from a TV show, a podcast, or a, what are we sick picked? I'd pick I'd pick some sanding paper. So, you can pick whatever you want. So, Dominic, Sarah, do you have sick picks for us today? What grit? Yeah. Oh, he's got he's got all the grits. So I'm just a wet sander. Oh, it's yes. Oh, I got, like, nine different grits. So it's for sanding three d prints. I I have a a little three d printed device that I sanded down, and it's so smooth. You can't even tell it's three d printed. It's beautiful.

I thought yours would be more about, like, dancing and,

break dancing. Node. Wes have tons of these. I I've I've probably sick picking for eight years now. Yes. Many, many, dance competitions. Absolutely. Yes.

I'm happy to go. Yeah. Yeah. Yeah. Go ahead. Like, one one thing I was, I I I love really, really good writing. And so one thing I would love to recommend to people is this this book I've been reading called the sense of style by Steven Pinker.

It's a book about writing, and it's it's a kind of a a style guide to the classical sense of writing, which is like its own its own style on its own. And it's read real written really, really well. I like to write, and I I try and, you know, be as good of a writer as I can. I love reading really persuasive, succinct, concise, impressive prose. And I think this book is filled with that, so I'd I'd recommend it, to people for sure.

Wes. Sam. I'm gonna feed this into my prompt, into my into my clanger so I can stop writing.

Yes. Yes.

I think mine JS, if you all like Minecraft or if your kids like Minecraft, there's a game called Vintage Story. It's a harder version of Minecraft, and it's sandbox, but it's very survival style. So you're the first human. You have to survive the winter. It has some, like, Lovecraftian horror vibes and elements, and you can mod it to your liking.

But it's really hard and really fun. So if you're, like, you know, into Minecraft and you want, like, to level up in some difficulty, me and my kids are just having so much fun playing that. Is it can you play multiplayer, or are you just taking turns? Oh, okay. Yeah. Yeah. It's actually really fun multiplayer because then everybody can, like, help with different tasks and things like that. You can do, like, a kind of collaborative effort.

My kids are both into Minecraft right now, so perfect perfect opportunity. Yes.

Nice. $24 too. I love I love that. You don't have to pay monthly for this thing. You just buy it once.

That's great. Scott sick of all the subscription things. Yeah. Me

too. I'm over it. It's too much.

Alright. Next thing we have is Shameless Plug. This is it gives you a chance to plug as many and whatever things you would like. Can you guys bring your plugs today?

Yeah. I would say, for me, mine's pretty boring. Just, feel free to follow me on Twitter or x, just at Dom Farolino.

Should be pretty easy to find. I tweet about random web platform API things that I'm working on and and stuff in specs and WebMCP. And so, I'd love to hear from people over there. I'm Sarah Edo on, X and other platforms.

One shameless shameless plug I do is that we're both speaking at AgentCon in San Jose.

That's run by the Agentic AI Foundation of Linux Foundation in October. So, they gave us a discount so you can get 25% discount with Sanity 25.

Wow. Awesome.

Thank you for, sharing that. We'll make sure that's all linked up too. Cool. Alright. Well, thank you both for coming on. Appreciate this. Let us know down below in the comments what you think of WebMCP, and we'll catch you later. Peace.

Thank you.