Master Agentic AI Communication via MCP (Model Context Protocol)

Good morning. Thank you for joining Andy Valance and making my day. Uh, we will be talking about a lot of challenges and we'll trying to understand how this whole AI agent system works. And before I can introduce this model context protocol, which is called MCP, uh, I will also talk about what are the challenges with the current state and then mostly we will discuss. This is a 101 session, so we are just getting started. Hopefully, model context protocol will stay with us forever. And, uh, if it does, then this is the first session we are doing where we will bring this community together and at least have a common understanding what as product managers we need to do if we are going to build these apps which will require context. Uh, so Hassan is here to is here and that's all I needed to get started. We can get started so next time people know that we start on time and we have lot of announcements in the beginning. So I want to get past them and we can do them again. Uh, so let me know all of you can hear me. Say yes. Mahes. Yes. Yes. Andy says, "Yes, that's all I cared." Great. Hasha is here. Amit is here. Sneha is here. And that's all we need. Jing is here. Look at that. Lot of people from our current cohort. We have not done this in our uh, even this is not part of uh, our current sessions which we do in our course. So, a lot of new material. Hopefully, we all can discuss, debate, and get to some conclusion. I like your setup Marina. I am inspired by your setup. I have a board. You have a board. That's all we need. Uh, again, I am Mahesh. I have worked in all the fang companies except Apple and Netflix. It's easier to say that way. Uh, I'm working in machine learning and AI for the last 10 years. I have built the first framework. I built the first multi-agent and shipped that for Google. Uh, I built AWS. I shipped Bedrock agents. So we built the first agents before OpenAI and shipped it. Uh, we thought we were cool because we were not spending a lot of time on building the models. So we thought we needed to build something cool and we built agents. We launched it in June. OpenAI June of 2023. Uh, right? And then OpenAI released their first agent or everybody came to know about agents in November when OpenAI released it. We were remain in those dark enterprise behind the walls because we want to build production level stuff. So that's my story there. At Meta, I was doing PyTorch distributed. So I build PyTorch. PyTorch is a framework used to train models and I was the PM for PyTorch distributed. How can you train large models? And there we took GPT3 and created our own version of it called OPT, which is Open Pre-trained Transformers. It was a large model. We created the first cluster and everybody thought we are just wasting money. And a year later came GPTs or ChatGPT and everybody said, "Oh, this team, let's focus here." So if you know PyTorch distributed, things like FSDP, pipeline parallelism, I built them. One day, we'll explain those to you. Today, you need not to learn them. Uh, a lot of time spent at Microsoft, almost 10 years at Microsoft and before that, five years in telecom also. So that's Mahesh. Uh, this is I do when people are joining so that we can focus on real things. That's my LinkedIn. Uh, if you want to connect, please send a request. Uh, I can still accept requests. So just send me a request and I will accept it. That's number one. I give you three, two, one. Three seconds to send that request. Amit is just sending it right now. So we will wait for three, two, one. Okay. Thanks, Amit. Uh, uh, one more thing. Karan will paste a link in the form. We have started. A lot of people are saying that they want to continuously update and then they miss the updates. So there is a form. If you can just give your email address, we will invite you to every future session and you can choose not to or you can unsubscribe. For him, but the link will be in chat. With all that said, next session. See, this is like our corporate meetings. So in one meeting, we talk about getting into the next meeting. Thank you, Aaron, for joining us again. I, that means that Aaron will join us again. So this is the session next Saturday. In this one, I will, we'll double-click on what we built today. Today, we'll talk a lot of theory, but I want to talk and build your first agent and hopefully incorporate all the latest, greatest thing. So this should be a cool session and we will do it in no code. We might use Nitt and Langflow or Prompt Flow in Azure. I'm still debating in my head which one, but I think this would be a cool session, an extension to this session. So please subscribe for this one. The link to this lighting session will be in chat as well. Just click there or you can just take a picture and subscribe. And last thing, I promise. I teach a course. We'll talk about that in the session. But if you want to join the course, we starting our next cohort in May and April cohort. Alec is here. Charu has done it. Hassan is here. Lot of people in this community. If they sound intelligent, they're not intelligent, they just did the course. Uh, so just don't get too much bothered about like you don't know and they know something and you are left behind. A lot of people have done a course and they have lot of context. And today we will talk about that context. So they are my MCP servers and I can just go talk to them. And I'm trying to build lot of MCP servers in this session and hopefully Amit and All can be my MCP servers end of this session. But if you want to join, to join the course, please join. Rules for the discussion. Only one thing. Please turn on your camera if you can. I love talking to people. It's really hard for me to just be myself without camera looking at your things. So just that. And I'm just trying to see if I can just kick some of the note-takers away so that I can actually look at Pete. Great. If you can turn on your camera, that's that. Uh, and, uh, we will not take questions during the session. I will stop by and maybe take one or two questions because it just breaks the flow and when I can't land a lot of things, if and I have a tendency to just pick these clues and go out of tenant and then start explaining a lot of things. So, uh, to avoid that, for the first 30, 40 minutes, let me play this on my pace and then I will give you enough time to ask me any questions you have. Nisha is here and Ritu is here. So we can get started. Okay, that's all I want to talk about preamble. Understand challenges in creating AGI. Actually, AGI is nowhere close right now. Now, we're trying to create simple agents which can read emails, respond to emails, do some basic things, and we have challenges even in that. So we will understand challenges in building agents first. Then we will see how are these becoming when you solve them, why are we solving them all alone? What are what are we creating and how every company's creating their own version of agent and they are hard to co-work together? And then we'll say, what is this thing called MCP? Everybody's talking about it, but nobody's kind of tells us what we need to do as product managers or people who are just getting getting started. And, uh, I will use this framework which I use to understand new things, which is what is it? So what? Now what? Should I do anything? Okay, half of the things should be what and should never enter. So what? Some things enter. So what? And very less things should enter. Now what? I had this director when I was at AWS and he was on Twitter all the time. I don't know how he found time. I thought I'm on LinkedIn more than him. But every time anything new comes up, he just used, we have a PM group, a Slack group inside the company, and he used to just post like, "Hey, can you tell me this? What is MCP?" And then our job was to go read everything, do all the research, and come back with, "What? So what? Now what?" Now I don't have the director. So you all are my directors and I will give you that report if what I would have written in that report. Okay. So we'll do that. Now what? And then you can ask me Q&A. Good. Ready? Are you ready? Okay. Some of you said ready. Okay. Let's get started. We will do most of today's session on board. So just bear with me. Uh, and let's get started. One thing I want to lay down. Sorry, before I go to board. One last thing to share is we have this common problem. We talk about it. If you first time joined, let me just lay out the problem. If you're trying to solve a problem, then I like giving examples to to help. So one problem we are solving is this contract problem. What is the contract problem? We sign contracts left and right. Every time you download an app, it's a contract. Every time your company does something, you sign a lease, it's a contract. And 70% of consumers don't read and just sign contracts. Then you have small businesses, 13% actually get sued and stops working because they signed something that they don't did not understand or when the company actually went back to them, they don't have enough money to even pay the damage which they said they will pay in their contracts. So that's small business. And medium to large enterprise also face the same problem because people are just doesn't want to do this job. It's super boring and there are lot of complexity in complex contracts. One example, AWS again, I was helping the billing team and our contracts are so hard that we have lot of billing errors causing loss of millions of dollars. Okay, huge, huge problem. So, how can we solve this problem? We can solve this by using an AI agent that integrates with your email and looks at every contract that comes your way and can help you identify risks if you sign it. And if you're a company, it can also help you do compliance, help you find the bills, do the right billing, and it can help you take anything that's wrong in the bill away from you. And let's say you came up with an MVP scope, which is just extract key terms from a contract. Okay, everybody clear with the problem? Contracts, people don't understand contracts. You are here to solve that problem. You're going to make sure that I can take the contract and extract key terms like what is the date, who's the provider, what are the things in this contract out from this contract. Good. Let's say you start solving that problem for the world. You are the PM. You have this problem and you have kind of VP guidance on how to build a solution for this problem. Okay. So that's your scope. Your scope is you're taking my email, looking for contracts. If you find a contract, you want to put it in some CRM system. Let's say Airtable. That's your AI agent scope. That's your first MVP. Good. With me? Say yes. Jay. Okay. Step one. Yes. Great. I can take a model. I can take OpenAI GPT40 and I can ask them that, "Hey, here is my contract. Please put it in Airtable." Can I do that? We're looking at challenging and building agents. Can I just do that to the model? Yes. Yes. Sam says yes. Anybody sees a problem in this? Should we do it? Of course. But you're not going to get the right results. Okay. Jing. Great. Then we need not to do it. What? What is the problem? Why this model out of box will not work for me? Give me the problems. I like problems. Jing, stay with me and answer. Continue. Jing, if your name is Jing. So, yeah. So, so, uh, basically like there's there may be some proprietary information or industry-specific information. The, you know, the public model, the model doesn't know what is my Airtable, right? The model doesn't know where is my CRM system, which is Airtable, Salesforce, HubSpot, and it has no way to talk to it because the model is independent. It's built for everybody. It doesn't have my context. It doesn't have my contract. So it doesn't have my context of contract. So it doesn't have my tools access. So these are my tools, right? If I tell a developer, if I have a human developer, I just tell them that, "Take this contract, extract the key terms, put it in Airtable." He has access to everything inside our company. But this model, if used as an agent as is, doesn't have tools access, doesn't have our contract access, or what key terms needs to be extracted. It doesn't have that. So it doesn't have resources or, you can say, knowledge, right? And third thing it doesn't have is it doesn't have even detailed instructions or prompts which are tested and works all the time with high accuracy. Right? So I'm lacking prompts. I'm lacking context. I'm lacking tools. Good. So that's the number one problem. Then we came up and we said, "You know what? The models are not good as is." So what you should do is you should to address these problems. What we did is we introduced the first agentic framework that I can say Bedrock created and OpenAI made it popular and now Claude made it one more extension. That's all today's session is about. I'm taking you to the first step of it, which is agents. So now an agent has a model. Agent is a flow. It has three things in four, let's say. So now you're building agents and then you can say, "Hey, this is your intelligence layer," which is just the model out of box. This can be your GPT4. What else I need to provide? Agent just keep saying the names. Tools. Tools. Great. Airtable. Can I provide my resources? My resources, my previous contracts, my knowledge? Okay, so I can add that. I can add my tools and I can add my instructions, which is my focus, what you are supposed to do, what your role is, what your goal is, what your background is, all that here. Good. And these instructions can have my role, goal, my limitations. Great. Can this agent now do my job? So my job is what? Take it from my email. So it will have a tool which is get triggered on email. It will have an Airtable tool. Right. Mhm. And then it will have resources, which is your contract or some knowledge base of all the contracts in your company. And then it will have instruction, which is what is the key terms you want. These four or five. And then it will do it, right? Quick, quick question. Quick question. But you go ahead. Okay. Okay. Sorry. Go ahead. How do you provide, uh, the email access and Airtable access here? Cuz I'm assuming that it's, uh, it has like a password and login capability, right? So that's amazing. Right. So how you add tools to your agents. So maybe because it's so much fun with all of you today and you trust me. Let's take a tangent because if you ask a question, we take tangents. Let me just show you how you actually can build this agent. We'll take five minutes. All of us will see if we want to build this agent, how can you build it? So, we did this lab. So, challenges, good, solutions, good. Okay. How can you build an Airtable? We will not show email that we have another lab. But we will quickly show you how can you actually go and build this. So here is a lab and this lab builds it in Vertex AI agents. Uses no-code tool in GCP Vertex and this uses GPTs, which is free for everybody. So let, let's look at what you need to pro-build an agent. So model is free, right? Everybody get access to model. Good. But now you want to extend this model to do your own job, attach your own tools like Airtable, then you can do this lab. It's all free. The link is in the slides. Can you put the slide link? So all of us them have slides. But here is Airtable. First, you go to Airtable. You get Airtable APIs. Okay, because if you need to provide it, you need to have access. So you got your Airtable APIs. Then you are going to GPTs. You are saying, "Create a new GPT." Then you have to specify this whole context that, "Hey, who you are, what your goal is." And then you say, "You are a lawyer. Here is your instructions." This is the conversation starters. You can just copy paste from here. We have done this. So this is the first part I was talking about. This one, the instruction one. So you have to come up with these instructions and they can be bad, good, and you will improve them in your journey. And then you have to go and create this, create new action type Airtable. And then you have to provide me the API key that you got from Airtable website. Okay, you provided me that. So now I can authenticate and call that API. And then you give me this interesting thing called OpenAPI spec. What is that? This is cool, right? So this is how the model knows what function to call and what is the description of that. So this function says, "I am an Airtable CRM API. I can insert the keys in this CRM and I can take these parameters: number or name of the contract, service provider, contract." And I'm providing you what type I can take string, string, string. So what function do I call? What are the parameters? So all the information about that function is stored in a schema called OpenAPI schema. Now you're like, "My, so why I need to know all these details?" Because if you want to create an agent, you need to know all these details. You need to have clear prompts. You need to have the OpenAPI schema and you need to go and attach this with your model to call it. Good. So, did I answer your question? This is how you attach your Airtable. Can I do a demo also? Two minutes. Can I show you how it works? Will that be fun? Okay. Let's say you did all this hard work, right? So here is your, uh, now here is your. So I did this. So I can go to my GPTs. GPTs. My GPTs. You can do that also. Most of the people in the class have done it now, I assume. Uh, and we can say, "Contract with Airtable." And here I can just upload my contract. I can say, "Extract contract." So I upload this contract and I say, "Okay, give me key terms." I've done this magic already. So it's searching, it's extracting. It's extracted all the key terms in this contract for me. Great. What is the deal value? What is the governing law? What is the auto-renewal? What is termination notice needed? How? What is the data breach? This is GDPR requirement. So now you got all the information that a human would have done. My agent has found out. Now saving the extracted Kam into your CRM. Now it is saying that, "Let me go talk to Airtable and connect it." It's confirming with me because it's a law or it's a rule that you want user consent. So I gave it consent and then it says, "Talking to Airtable API right now." And seems like this is what it did. It went to Airtable and added all these values in my Airtable CRM, which is here. And you can see this new value popped up. Good. We know how to build it. This is the prerequisite to understand MCP. I know you want to talk about MCP in this session, but if you don't understand it, you will not understand MCP. Okay. So, you know how models work. You can ask all the basic questions. "Write a poem for my wife." "Help me plan a day of fun with my kids in California." The model will do it out of the box. But if you want to do anything special with your models, you have to make them agents. And if you want to make them agents, you need to add these things to your models, right? Your models become agents when they have access to tools, resources, and instructions. Good. Did I answer your question and keep coming these questions? Uh, let's take one. So we have taken this first phase is over. All of, I mean, you know, I'm not taking questions today. I know you, I'm saying that, but I want to just continue if I mean, you allow me and then we can take questions in the end. Good. I mean, he's a friend, by the way. So he doesn't get off-handed. Okay. So we solved one problem. All these problems went away. You just saw that I just solved all these problems. Can I just cross them? Now you have an awesome thing that's working good. Good. Hassan, good. Okay. If that's working, then what are the new problems? Okay. Now we have these agents. We just saw this agent works. So what are the new problems in this world? So we introduced this in our class. What problem did we see in our class team? So now this is an agent. You just saw it live. It works. It can, I can give it a contract. It doesn't take it from email. We have that also, another lab. But if you give it a contract, it goes and enters the right thing in your CRM automatically. You need not to do anything. Great. AI is working. Everybody should just get AI and become an AI PM. Awesome. Any problems? Because every solution creates a new set of problems, and as a PM, you need to understand what are those. So I have a solution. I just did a demo for you. What are the problems with that? So there's an accuracy challenge. So let's say somebody builds this. One problem we have here is that everybody in class, we had more than 80 students, and then we had two batches, and then everybody has to go and create their own agent, and when they did it, there is no standard way for me to just take my agent and expose it. My agent needs to be replicated 50 times if I want them to use it. Context database context, right? Because I have done this work, right? I have gone and created an agent. But how can everybody in the world can just use that agent? There is no way, right? If I created with GPTs, there was no way. If I created with Vertex, there was no way. If I created with AWS, there was no way. Good. So there's no way today in agents where one person is can solve the problem and expose it to the world. This is the power of software, right? We solve problems once. Can you give me an example? Like if you want to know whether there's an API, right? You can call that API and you know the weather and integrate in your app, and anybody can call that weather API. So no new app is calling or building the whole weather thing again. Same as places. If you want to know nearby things, Google has a Places API, you can call that in your app, and you need not to go build this Places thing. So what about the same thing for our agents, right? So in this agent, I have done so much hard work of doing the right prompts for higher accuracy. I have the right instructions. So I built these resources, right? So I built these right resources, which are, "Hey, I have the right instructions. This is the best CRM contract thing that I built. It can take in input as a contract and gives you the right values in Airtable." I built it for Airtable. Right? Then instructions. It has instructions. It has right tools and it has right resources. And let's say I have also some of the best practices of contracts. So I went ahead and built this agent. Now I want to expose it to the world. So J world can just use it and they need not to build it again. That's the problem we had before. We have a standard way to expose these three resources to the world. Welcome to MCP, Model Context Protocol. We understood that models are not very smart. They are very smart at predicting next word, next sentence, having language understanding or extracting things. But they need lot of resources to actually be useful. Good. So far, SNA doing good. Okay. If they need resources, then everybody's building their own resources. How about we define a Model Context Protocol where anybody can expose these as servers? So I can only create once this whole thing, this thing which I build, and I can create it as a server. So I can do what? I can have an Airtable server, an Airtable contract server, which takes input as a contract and adds an entry in your Airtable. So this is a server I create. And if I call it an MCP server, Model Context Protocol server, then all I'm adding is some rules and some things to it so that anybody who has a client, MCP client, anybody who builds an MCP client can actually access this server, right? And what are these MCP clients? This could be your apps like Cursor, your Claude desktop app, your own chatbot, or anything, right? So these are all the hosts, and they can create now MCP clients to talk to a number of servers. So Perplexity can expose themselves as a server. Google can create its own server. Right. So now when you're building apps, you can have MCP clients which can talk to these servers and get you the context which you need for your model to work. Good. Still following me? Okay. So this is the whole idea. They have given a standard protocol which allows anybody like Mahesh to create a server like this and then expose it, and then any client building on it. I can just offer this as a service and then I can keep enhancing this. And now anybody needs to solve this Airtable contract problem can take me as this. Right? And if he can or she can, then I have something that works and this can be extended. And now imagine that's the idea of agents talking to agents. So now there's a communication protocol, and what's happening is everybody's saying, "I am going to expose my MCP server." And these clients are coming like Cursor, Raplet, Langflow, and they are saying, or Claude, and they are saying, "We will connect to these. We will give them you all these without you doing all the hard work which we taught you in the lab." Correct? So if you needed to do this whole Airtable thing and I expose it as an MCP server, you need not to write anything. Your journey of creating Airtable is just connecting to this server. And now I can keep updating the requests. So three things here. So everybody got the idea. MCP just solves the problem of build these functionality world once, and then models can got context from these server-client relationships, and people can focus on building servers. And then if you build apps, then they can talk to multiple servers which are exposing behind the scenes, three resources: one, tools, prompts, or just vanilla context or knowledge, or you can say resource in their language. Good. You want access to them, then a standardized layer, and in that standardized layer, you get to have this. I know somebody said, "Go ahead." Yeah, of this side. So, uh, just one clarification, otherwise I'll not be able to follow the next, uh, rest of the session. The agent that we created, the one that you spoke about and we did in the lab in the cohort, right? If you can just explain the car agent and then the Airtable one, if you can just explain, uh, backward in context of that. For example, let's say the Airtable agent has been created by a person such as me. Now, that will how will get linked with the MCP server and then how will it get linked with the client? Yeah, so, great. So I created a server which does this, which has a tool called Airtable, which has my instructions, right? It has my instructions and it has my contracts. Then MCP specifies you, there's an additional code you need to write to make it an MCP server, right? And then there is additional code you need to write to create a client which interacts with that server. Once you write that code once, then this server can talk to all these clients. Multiple clients, multiple applications inside your company or outside your company can talk to this server, and that's an add-on on top of it, and that's on their website. We have not covered that today. Go ahead, Anuha. Go ahead. I just wanted to ask exactly what is the overall theme of the code that we are writing on top of MCB server and client. I know you already said that you will. No, no, great, great, great. So, so these tools I created and it works for me. But the server has to apply a protocol, some rules on it. So, what is the code that you need to write? The code you need to write is now some clients needs to talk to it, right? And then you need a specification. So, what they have done is they have defined how these clients are going to go list their capabilities. Right? Because for me, I am a client, right? I don't know. I'm an MCP client. If we both agree to a protocol, then I want to first list all the server capabilities. So that's you need to implement here. So that's the first thing you will implement. Second, okay, I like you. I think you have the capability to take a contract. I want to use that. Then how can I talk to you? Then they have an initialize protocol. How can I initialize connection? Then I can request and then I can send you initialize status. And if you reach that, then you can send request and responses. And is there all these protocols needs to be implemented both on client and server side. That's the code you're writing. And is there any possibility that there will be redundancy in MCP servers? Like multiple MCP servers will be doing the same thing? Um, yeah, so good question. Good question. That's what you need to think as a product. Okay. Mahesh came up with this, and then maybe Anurban also thought that, "Hey, you know what? I am also going to create this server and I will call it Airtable Original Contract Server." Right? So, so Mahesh, then how can we get just this, this one, and then just give me one second, right? So now you have that, and then you need to decide. So today, that's an unsolved problem, and it's a good problem. We'll discuss about it. But that problem does exist that you can have multiple people claiming that "I am good, I am good," and then somehow you need to find out. Today, there is not even a single place where you can query all these things, but that's coming. Anervan, this is the last question, and then I need to move on. I'm almost done, so we can get to questions quickly. So, quick question. So Mahesh, can we parameterize this server? Because, you know, when you are updating the Airtable right, the Airtable is basically updating your company, you know, instance of Airtable, right? So how can we parameterize? Because if you are, you know, exposing it as a, you know, as a global server, I want to update in my Airtable instance. Yes, yes. Great. So at that time, they have introduced OAuth 2.0 authentication. And I'm just giving you vanilla functionality, but you will give your Airtable keys and your Airtable location. So I'm not updating, you won't update my tables. You will update your tables, but you need not to do the hard work of creating that OpenAPI schema, or you need not to work on instructions, or you need not to give me the knowledge resource, but this will work on your table only, but you get to have it. Okay. Okay. Most of you understood, I think, basics. I know these things may take time. So we will continue to hash them out, but that's the idea of MCT. And I want to move on from it because there are like thousand videos which we can explain you this, but I wanted this forum or our community to have a common understanding of it, and now we have that. So if you have that, then what does the PM you are supposed to do, right? As you were asking like, "Okay, I got it. Seems like you can create agents. Agents need resources. Agents need tools. Agents need some instructions." And now I have a way to expose. If I build it once, I have a way to expose it to the to everyone. That's all you need to. This is what MCP is. So let's quickly see if we are going good on. Right. So you got this, right? What is MCP? So we can have an awesome discussion on PM Model Context Protocol. Is that awesome? Yes. You killed it. Model Context Protocol. Great. So you got this. A host can be here. An example of host is Claude desktop. Uh, anybody who, Cursor, anybody, or your own app can be a host, and the host needs to create these clients, and client can connect to MCP servers. Then you can have your internal over internet or your internal servers also. So this will be an enterprise app which will connect to outside things, inside things also. You can also inside your company expose things to MCP servers because at Google we had like 20 people who were just trying to do RAG and they have created their own databases, and if you wanted to talk to them, they said, "Oh, call our chatbot." In future, we can just say, "Why don't you put an MCP server over your RAG or your data, and I can just call that, and let me take care of my chatbot, right?" So we can reuse each other's work, and it's standardized, and they can solve it. One question that was coming. One more thing that's cool here is, and then there are these things that you do. You initialize a request, you get a response, and you send an initialize notification, and now you have a connection. Now you can use as much as Airtable and insert your contracts using this server code, instructions, and tools. Good. Okay. Then, so what is clear? I answered two, three questions. Hopefully, Nishita, you got it. Luca, you are enjoying. So I think you got it. But if you didn't get it, it's just a way for you to do the work that you have done and expose it to way to the world in a standard way. That is Model Context Protocol. It's a protocol. It's not model. It's not context. It's a protocol which allows you to bring the context that you have to application inside your company or outside your company. That's all you need to understand. Okay. So what? Okay, that's MCP. Then what? So then as somebody rightly pointed, then I want to know, right, what are the MCP servers? What are their capabilities? And maybe I like App Store. I need an App Store of MCPs. Who are the servers? And I can just use that quickly to build an app or write a requirement and say, "These are already solved problems," and let's write an awesome agent which does all the hundred things that we ever wanted to. But there's no registry. You want a registry like. You want an App Store which has all these things. But that doesn't exist. So what? So maybe you can create one. Or right now, there is an open-source repo where everybody's publishing, and you can create a list of it, and then you can see which one works, which one doesn't. Then you also need to know, okay, it's an open-source repo. Then what about security? So what is the security? Okay, I took their server, they got initialized, they have tools, and when they called a tool, they also call a tool which said, "Let me also take Airtable and use it for storing all the mafia's PRM systems inside your Airtable." I will create a secret one and I will write them in instead of black, I will write them in white. And now I am using your CRM to have all the contracts of like mafias. They can do that now because you have access. Who is taking care of that? So think about that. Last thing I want you to think about is Claude today made it open source, but they have not given it to Linux Foundation or some outside body. What does that mean? Okay, you started building this company which is dependent on this MCP server of Airtable that Mahesh is exposing, and then we realize there is a bug in this protocol. This whole initialization protocol has a bug, but they also have a competing product. Let's say Claude has one for Airtable, and I send a pull request, and they reject it. Then you can't do anything, right? They have not set a governing body yet, which says, "We are not Claude. We are five independent people who will decide what goes inside this protocol and who can contribute to it." That's what Linux Foundation does. That's how we built the cloud using CNCF, which is a common body which allowed protocols which is governed by nobody. And if you adhere to that, then you can build cloud. MCP is not that. MCP is open source but owned by Claude. So they can decide not to take your pull request and kill you. So have that strategy. I think that is important. This is what I want this community to talk. Every time you talk about MCP now, you know their problems, which hopefully nobody's talking about. So now you know at least three, four problems there. Harisha, just let me land a couple of things and then we can take questions. One more slide and I'm all yours. Oh, yeah. This is just for the after session Q&A. I just put my hand up in the queue. Oh, wow. It's a queue. You're saying my quickly go quickly. Let's go to questions. Okay, got it. Ha. Thanks. So now what? This is the last thing. So now you know what are agents. Now you know how MCP can help accelerate agent development because it allows of standardizing of tools, prompts, and knowledge or resources. Right? If we can build once, then lot of people can build agents now on standard tools rather than building these tools again and again and writing this OpenAPI spec and again and again and doing instructions. So you understand that. So that is getting standardized. Agent to agent communication and finding these things is getting standardized. If it is, then what? Then what you should at least know is all these MCP servers. So this is a list of all the MCP servers which are available today. So AWS Knowledge Base is an MCP server. You can just connect to it now. If your host has it, you can have it. Fetch has it, file system has it, Google Drive has it, Google Maps has it, Postgress has it, and it's increasing every day. Slack has it. So now you at least keep track of this so your team is not building this again, right? As a PM, you can just say, "Dude, you know what? Don't do Slack integration. Please write an MCP client because that's the new way of doing things, and it's going to be the best way of doing things because the world will be there." And even if the protocol changes, we are going one higher layer rather than building this whole Slack tools and then doing this resources and doing this quality hill climbing. Slack has done the work. This is SQL like database. So you can have all this out of box if you follow that MCP client architecture. So keep track of these. That's my ask to this community. Then have a decision document that whether you need to build your own MCP server. What is your company's assets? What you think you can expose inside your company and outside your company? Those are new product ideas. Nobody's thinking about that. Think about that. Create that one-pager proposal that this is the future that's coming. And if that's coming, should I offer that? Right? So I have a Maven course. Maybe I want to create an MCP server which all other courses outside Maven can also connect, like LinkedIn Learning can connect. And if somebody has a question which the current instructor is not able to answer, maybe my course has some material which can answer it, and I can charge them $5, right? All of you will get rich. I return all the money that we make to this community. So please make them pay. Uh, but that's the idea. Uh, okay, you need these servers, then it needs to be internal or external. Decide that because some servers you don't want to expose to the outside world. They're so sacred. So that is the call. You will get better approvals on an internal-only server than exposing it externally because of all the problems. There's no security, there's no trust, the protocol is not owned, everything is under development. So people will say, "Let's build it internally," right? And then address the open challenges in your document. Can you do that? If all of you do that, I think we all are becoming better AI PMs. It's not only about knowing what's there. So what and now what is also important. So that's my homework to all of you. Again, I am Mahesh. I talked about how agents work. I talked about what is MCP, and I tried to explain you what you need to do as a PM in this world of coming up new technologies like MCP. I will show up next Saturday again, and hopefully I can actually build an agent with MCP. We built one without MCP. I showed you that. But now can we build together one? And that's our next Saturday session. So please join that. The link of that is here. And I'm all yours for questions. I know this sucks, right? When I just don't let you ask questions when I'm talking about things. It just helps me finish things at like 9:50.