MCP is the key to an agentic internet | Humans Talking Agents Episode 6

Key Takeaways

What an agent is is an agent has agency and agents get agency through you releasing your agency. Say that fast five times. That would be hard. You are allowing agents to make decisions on your behalf. You're giving them your agency. It completely changes the way that businesses do business, right? Because you're no longer marketing to the end consumer. You're marketing to their agent. People, in this case, agents interact. So in the same way that like companies create and expose APIs with like standard formats for people to interact with now I think that every company will add on a second thing which is a server that can access these APIs. So I think there'll be probably a lot of marketing from like a lot of these individual companies towards uh agent creators to say hey prioritize my server or I'll pay you to include my server within your agent so it accesses Expedia versus someone else. What's in between? If that is zed and we are at A, all the letters in between are unclear. It could go a million different directions. I just feel pretty strongly that's probably the end game. Welcome back to another episode of Humans Talking Agents. Now, if you have been on Twitter or I guess now X uh anytime in the past couple weeks, you have seen MCP. It is all over the internet and it's a brand new way to build aic tools and agents. And so that's what we're going to be chatting about today. How can you use it? What is it? and what's the future of this sort of uh MCP framework? But for the uninitiated, Daniel, what does MCP stand for and what is it? Yeah, so MCP is model context protocol. Now, those words mean nothing. Uh, and I'm going to explain generally what it actually is. So, by the end of this video, you can go and explain MCP to your friends. Uh, and then, you know, make tweets and actually know kind of what you're talking about. So, um, to get started, uh, like right off the bat, none of this is very new. So if you remember tool calling or function calling um that GPT came out with and then everyone else adopted that's kind of the predecessor to model context protocol. And so function calling or tool calling was basically you give your LLM a list of tools that it can use. So for example, web search or retrieve from Google sheet. Um and uh you give it a little description for it and then when a user is talking to it, the LLM can decide when it needs to use a tool like web search or retrieve from Google Sheets. You also give the LLM a list of requirements. So if you want to use this tool, you got to ask the user for their name, email, phone number, whatever it might be. So the LLM is able to say, "Cool, I probably need to use this web search or retrieve from Google Sheets tool, and to use it, I need to ask the user for their name." uh you know phone number, email, etc. So that's tool calling and then basically what it would do is the the AI would basically send a uh like a a payload like a little JavaScript step uh that says uh here's a tool name, here's all the requirements. Then you would take that on your end and make an API call with it and put information back to the large language model. So that was the first thing that came out. But the problem with that is that every model had their own way of doing tool calling or function calling which made it kind of annoying if you were working across these models. On top of that, you as an individual had to like write all of your own code to then say, okay, once I call the tool, like what am I going to do with this? And so that's where anthropic came out with MCP or model context protocol. And basically, it's a guideline of how um a lot of models can use tools. And it's actually a bit better than tool calling because it's easier to be able to create and reuse across models. And on top of that, companies can create their own tools or their own servers in this case that'll allow you to access their APIs. So let's actually share screen a bit and walk through what a real MCP server looks like. So here we are on the MCP or model pro model context protocol website. And over here is kind of a little architecture diagram that they're showing where you've got your host. So imagine this host is your large language model. It's able to look at a couple different servers and then use those servers to call these tools. Now this seems kind of abstract. So let's actually go and look into one. Uh for example, so if we go to example servers here and we're actually going to go ahead and look at GitHub. Um and so GitHub uh is you know where you can uh store your code, create branches, etc. So this is the GitHub MCP server and you can see right off the bat here, here are the tools that I was talking about. So there's a server for GitHub and here's all the things you can do with GitHub. You can create or update a file, you can push files, you can search repositories, create one, etc. And with each of these, you can see that here are all the inputs that are required. So very very similar to function calling. So what's happening is when you're using your large language model, you basically give it access to this server that already has all of these tools that someone else has created. So without you needing to write all of this yourself, you're able to connect to it and LLM can say, "Okay, cool. Here's all the stuff that I can do with GitHub." So what's happening now is as you talk to your LLM, it's functioning the exact same way that tool calling was functioning where it's determining when it needs to use some of these tools. So push file, creator update file, search repositories, and then it's going and collecting all the information. The next part is like, okay, cool. So I've got all the information. Now I actually need to go and make an API call to GitHub to do the thing. And that's also what the server has access to. So if you go in the operations file here, you can see that for search, for example, u you know, here's the code that once it actually has requirements, it's going to go ahead and actually just straight up use the GitHub API to search through issues and then return that information back to the model. So it's very very similar to tool calling if you're already familiar with it. It's just an easier packaged way because what this does is um like someone can create a server for GitHub and then they can share that with anyone. And now any model that I'm using, I can use this and access all the tools that someone else has already created. And what's even better is that companies can create their own servers as well. So GitHub can create their own server. Yahoo Finance can create their own server. So if you're building a app or an agent that needs to access stock data, you don't have to go and write all this yourself. Yahoo Finance may already have an MCP server that you can connect to. They've already got all the tool calls that are available. And so you can already start using their API without writing a lot of that code yourself. And so that's really what MCP is. And so the real power is in that you're giving everyone the ability to kind of create these servers for their APIs that anyone can access. So it makes it way easier for you building an agent to access a service. And as a company, it makes it way easier for me to be able to let you use my APIs. And so that's really all MCP is. So there's nothing like totally crazy new that's happening here. It's kind of like just a better way of of doing this across all these different models. Yeah, there's a few interesting things with MCP. I mean, we made a joke at the start of like you you don't have to think about the, you know, what it's actually called. I think by clearly calling it out as a protocol, though they're trying to make this the standard for how the internet interacts with agents, right? Um, so it's actually I think that's it's a really smart call out of like it's not just the functionality, it's the branding around the functionality. And by calling it a protocol, they could have the chance to make this like the standard for how you have like an agentic internet, right? I think the second thing that's interesting is like the core concept of MCP is not new. In fact, this has been around. So for those who don't know, voiceless started in like 2019 with like the Alexa days essentially, right? And the core idea of Alexa was actually very similar to MCP. I think this has really been the dream of everyone who's been in the assistance now the agent space forever which is um I can ask an agent to do something for me and it is able to orchestrate uh complex tasks and it can also orchestrate across multiple different vendors and APIs and things. So now as you start to um uh if companies start to rapple their AP uh APIs with an MCP server what becomes really interesting is that real like that dream is potentially going to become a reality. I could go to my personal assistant and say, "Hey, uh, I want to go to Paris, right?" Um, and it has a KB has a knowledge base of all my, you know, past travel, my past preferences, and it goes, "Okay, well, what types of hotels does Braden like? I'm going to go to Expedia, right? I'm going to go use their MCPc per uh MCP server to get the right hotel." Um, and it has all my preferences. So, boom, it's going to get the right one. Then it goes, "Okay, well, how is it going to get there? Oh, I got to go to, you know, United Airlines. I got to book a flight. And so all these different places or all these different like destinations on the internet, instead of you, the user having to go and use the front ends of um these brands, you could actually just have an agent triage uh effectively and make that have all the decision-m process for you using these MPCP servers. Um now, there are a whole bunch of implications that come out of that. First of all, like does that allow for uh you to have enough agency as a consumer? What's the you know, do websites matter anymore if you've MCP server? There's so many different things that come out of this idea that's been around forever. Forever being the context of AI assistance, but that is now quickly becoming a reality, probably faster than people thought. Um because, you know, I don't think we thought we were going to have a standardized protocol. Um we don't know if this is going to become the standard. You know, there might be multiple standards. Who knows? But the fact that we're moving in that direction starts to beg some of these questions of like, do you need a website? Uh, are we all just going to have personal assistance? Like, what does brand mean anymore in an agentic world? Bunch of big questions. Yeah. I mean, I guess like at the end of the day, really, I think for me, all this is is uh just a easier way for for agents to be able to access APIs. And in the same way that like an API allows you to access someone's like server or someone's database or someone's technology um like this model context protocol is basically a way for agents to be able to access APIs um in a very easy way. I I I agree. I think that's the technically yes. Right. What an agent is is an agent has agency and agents get agency through you releasing your agency. Say that fast five times. That would be hard. Um, no, but like the whole concept of if you are allowing agents to make decisions on your behalf, you're giving them your agency, it completely changes the way that businesses do business, right? Because you're no longer marketing to the end consumer, you're marketing to their agent. Oh, that's actually really interesting. Yeah, that's a that's a really interesting point. It's like kind of the same way with like um how search is actually changing with like Perplexity and GPT, right? Because you can't just like do a bunch of SEO stuff now to get surfaced on Google's engines. Um, and I'm sure at some point there'll be some sort of bidding system for articles and content uh to be able to be better ranked within like training data or with like a web search for example. I mean, it'll probably end up being a whole new type of like, you know, I don't even know if we call it SEO anymore, but like you're optimizing your services, your MCP wrapper to be surfaces the right result for that particular user. Yeah. Right. Um, which is, you know, like things like billboards, things that are in the real world, there's no point because I'm just going to ask my agent, right? Yeah, it is interesting because I think especially when you have a lot of these like um search services that all compete with each other. So, you mentioned Expedia, Booking.com, Travel Velocity, like all of these ones that Kayak, they all do the exact same thing, right? And they're all super SEO heavy heavy today. That's the core of their business. But it's kind of interesting because they're like they're they're like flight aggregators and so or or travel aggregators. And so yeah, like models will basically well it'll be the it'll likely be the uh agent creator that'll determine because you have to give it access to which like services you offer. Yeah. So I think there'll be probably a lot of marketing from like a lot of these individual companies towards uh agent creators to say hey prioritize my server or I'll pay you to include my server within your agent so it accesses Expedia versus someone else. There's a whole like there's a whole aspect there. Yeah. I mean we can go into a rabbit hole but there's so many different business models that completely change because the value of an aggregator like how they make money is they get referral fees right now could you have an agent that itself collects the referral fees like dis you know and discounts it like do the agents become the aggregators do the agents choose the aggregators do the direct brands just at this point realize they don't need aggregators like I think the probably the simplest model because like the the a lot of these aggregators like Expedia right we'll just take them for example they do a lot of uh really important work because all the airlines that they're working with like all have their own like formats of data etc right so half the work on like an Expedia is working with all these providers to make sure there's like a standard of data that they can import or cleaning it up or modifying it themselves so they can actually surface flights from all these different airline providers um and so you know it's still very valuable that they do that because now your agent just has to interact with one service instead of like 10,000 services and so that part I think is still valuable and I think then at that point Like I still think it becomes kind of like a bidding system, right? Like if you're like Apple and you're building the next version of Siri that's hopefully good. Um you're probably going to be doing brand partnerships with a lot of these companies to be able to book travel services via Expedia or whatever, right? And so I think there's like obviously the big companies that will have these like uh more general purpose assistance that they'll probably have brand partnerships. Yeah. um or even a whole new partnerships team with all of these services to figure out who they use and what they use and then a whole community aspect as well. And then the other side will be what about like for I guess all this smaller kind of like agent creators too, right? Like there might be something on that side where if your agent is getting enough traction, you might be able to also kind of reach out and almost do like sponsorship type sales um for your agent to use a specific service. You know what's interesting about so you know we're talking a lot about the MCP servers and how they would compete for you know being able to fulfill these tasks essentially right like MCP is essentially being like the modern website like SEO but if you look at the other side of it if you then view the agents as the browsers all the browsers today seem focused on traditional search and like they're trying to discover the internet in a more efficient fashion for you. So perplex perplexity is a good example, but in the world that we're talking about, it would actually make more sense to flip it the other way around. And the agent should spend more time learning about you so that it can actually find like if we assume that the um the problem is not just about simply indexing and finding all the right resources, then the thing that actually makes the best decision for your agent to get you the the correct MCP server would be to learn more about you, right? Yeah. And so like the the browser should be spending more, you know, it's funny. People complain all the time about Google data mining you and and stuff like that. But um you might see a proliferation of new startups and new companies. Heck, people might be able to build their own browser in this world if the browser's, you know, an agent, right? Like we might see a consolidation of a single couple large agent browser companies, which are essentially your way to portal through the internet and get things done that you want. Or you might see people be able to build their own like it's going to be the wild west. But I do feel like they're not focusing enough on learning about the user and learning their preferences for a world of agency where the agents are making decisions on your behalf. It's not just about finding what's available. It's what's finding how you yourself would actually choose to make that decision. But I think that will happen cuz like if you think about it from a from an agent standpoint, there's the the larger the larger tech companies will have more of these like general agents, right? That's what they're all trying to do right now. Then you're going to have a whole bunch of like per industry all of these smaller more specialized agents. So for travel like I can imagine that there are probably multiple companies that have travel agents including these companies like Expedia etc that use their own services and then there's going to be in the same way that there's today there's aggregator sites and then there's aggregator of aggregator sites that search like like kayak right it like searches expedia and all these other places let let me ask a dumb question let's take for example uh Siri yeah as let's imagine Siri becomes good and Siri becomes an agent that knows everything about you, but you know the it's encrypted, it's on device, you're not worried about privacy. Okay, let's put that aside for a sec. Would you like which would give you the best results? The Expedia agent which doesn't know anything about you, you use it one time or the thing that knows everything about you. I would actually argue that you almost this might lead to further consolidation of services and you don't need tons and tons of different agents. you just have a single agent that is your personal agent that's interacting with the internet through MCP servers. I think you can say that today. So there there's a couple pieces to this, right? One of them is understanding about you, right? Like you are Braden, you travel frequently, you like to stay in like whatever certain accommodations, cheap hotels investors listening. Um and uh that's like one aspect of it. Then the then I think there's two other aspects. The second one is obviously like the experience like the UX of actually going and doing this. And I still think that's important, right? Because at the end of the day, like I think people will say like, "Oh yeah, I'll have this agent go and just like book all this stuff for me without my approval." I don't think we're there for a long time. I think you're still going to have some sort of interface where you're still trying to figure out cuz like it's not just like you got to figure out flights, you got to figure out timings, you got to figure out all this different stuff. So I think there there is a UX component to this. And then the third part is actual like access to data, right? And like access to data is through a lot of these individual companies that have access to different data. And so, you know, your Apple one will probably have a big brand deal with Expedia and be searching Expedia, right? But I want something that's going to search like all of these like random booking sites that I can do, right? You're probably going to have something specialized for that, at least in the short term. This is the thing. I think it's actually quite easy to see the endgame. Yeah, it's really hard to know the steps in between. And I think that's kind of we're almost debating sequencing at this point. Like I think the clear endgame is I will have a personal assistant and I say hey I you know this is almost how EAs work for like high net worth individuals right like that EA has learned their preferences over time working with them they know their travel schedule they know what needs to happen it all just gets done right makes your life super easy that'll get democratized out when we all personal assistants and you're able to go and you know uh get everything you need done to make your life just as easy as possible. Yeah. um what's in between. If that is zed and we are at A, all the letters in between are unclear. It could go a million different directions. I just feel pretty strongly that's probably the endgame. Yeah. I and I I think again like we can look at existing patterns where you know if if there's like an 8020, right? 80% of travel is like a certain part of things and 20% of travel is like you know the more niche I don't know. I don't know a good example. But um I think that 80% will 100% be chewed up by these large companies even today, right? Like I go on Google flights, I look at a flight. I book a flight to Google flights. It sends an email confirmation. Google figures out when that flight is departing. It puts it in my calendar. That's basically my own like as executive assistant. The other thing is like I'm going and buying the flight, right? And so that's already happening today. But if I'm like trying to plan out something that's really big with like like I'm doing bachelor party soon, right? There's a ton of people involved. Like I got to figure out like all these logistics, talk to all these people. Um like I still need to do a lot of work on that side. Um and like even then I'm using a bunch of like smaller specialized services. Yeah. Yeah. Um, so I think in the short term we'll probably find the same pattern where like 80% of these tasks are what the large companies will do. But I still think we're forever going to have this 20% that's so niche doesn't make sense for the big companies to focus on it that you'll have like these smaller more specialized whether it's agents, whether it's some sort of like agent UX hybrid, I don't know what it is, um, that's better geared towards that like individual task. Yeah. I think what's kind of going back to what we chatted about at the beginning like I think what's interesting about MCP is not actually the capabilities it's the it's the promise of a standardized set of capabilities. Yeah. Right. Like the fact that every company when when they think about like how do I enable agents to interact with my service? I think before it was maybe a bit more of a bespoke thought pattern but now it's like we can just set up ancest it's just brought it top of mind it's like it's cool because it's like a new it's like a new new set of documentation or standardized documentation that allows people in this case agents to interact so in the same way that like companies create and expose APIs with like standard formats for people to interact with now I think that every company will add on a second thing which is a server that can access these APIs y and like that itself is really exciting because it it it kind of like opens up and democratizes access to these APIs. And I think that UX of connecting agent to these servers and where these servers are hosted will become like better and better and better. So it's kind of like agents having their like Zap year moment, you know, where it's like there now becomes a really easy way to connect to like thousands and thousands of end apps directly. Yeah. My my brain just goes to this g be a random tangent. I feel like my brain goes to um for all the big services it makes sense and then I was think about voice flow. Okay, what do we do for an MCP server? But we're an agent building company ourselves. If we enabled other people to if we enabled agents to set up their own agents with their own like it would just it's just I mean we're kind of already doing that, aren't we? With some of the stuff we're talking exception of how many layers of agents and well it's it's more like the co-pilot discussion, right? It's like being able to have something that can access our own internal endpoints to make the building experience easier to autoimprove your agent to go through transcripts. All of these things that you know we may or may not potentially do. Yeah. Yeah. Yeah. Road map unconfirmed. Um but any that's very very cool. Cool. Any I mean there's a lot of good stuff. I think there's a lot of good stuff in there. I think that's a solid way to maybe to wrap it off. Um but yeah, let us know what you think in the comments. Um curious if that explanation helped you understand what MCP is. that like you know took me months and then I was like let me actually go read on this and I was like this is way simpler than I thought. So uh yeah let us know in the comments if this made sense or if there's any other concepts you want us to explain. Thanks for tuning in. Bye. [Music]