MCP explained: What everyone should know about Model Context Protocol | Amazon Q Podcast

Key Takeaways

Welcome everybody to the Amazon Q podcast. I'm Chris Shay. I'm a solutions architect at Amazon Web Services and I'm joined today by Trevor Spires, a senior solutions architect at Amazon Web Services. Today we're going to be talking at depth about model context protocol, MCP. Trevor, it's so great to have you on the show. I really appreciate you uh making time for this. >> Oh, of course, Chris. I've been looking forward to it. This is like one of my favorite topics, MCP. So, thanks for having me on. >> Let's jump right into it. MCP, model context protocol. What is it, Trevor? >> Sure. I mean, we can break it down to its to its three words. Model context protocol. So, it is a protocol to enrich contexts of large language models. So, uh the the analogy I like is from the creators of MCP Entropic who uh define the protocol last year. uh they they describe it as a USBC port for large language models, but people should think of it as just an easy standardized way to plug any tool into a large language model to extend its capabilities. So people may have heard about agents. It's kind of like uh agents but uh a standard way to make models do things that you want them to do and interface with tools you want them to interface with. I think what's so exciting about that too is the fact that we're really seeing a huge shift and I think it's related very heavily to agentic AI. How you don't just have a prompt-based text output from an LLM that you have to then you know paste into a different format or work with like you can actually have these you know with the power of MCP have AI models take action on your behalf and know how to in a standardized format integrate with other tools and resources and I think that's huge. >> Absolutely. Yeah. I I I've been having this conversation with customers a lot lately of around agents and MCP and the discussion is like you know we often are familiar with the idea of going to a chatbot or an AI and asking a question and having it turn a response but uh what if you could build something that it it gives your user the answer before they have to ask for it. Right? So imagine like logging in at the beginning of the day and having like a briefing of all of your all of your meetings and all of your important um to-dos for the day. Or another example, I work with a customer that recently uh we we co-authored a solution with to uh send stock digests to customers effectively, proactively on demand. And that that's got me thinking a lot more of uh asynchronous agents, agents uh with MCP potentially that are just working for us behind the scenes. And that that's like a whole other level of productivity I think folks don't always think about when they're stuck in copy paste chatbot mode. >> Yeah. And I've seen that myself too. Like I think a great example too of what we're seeing here and how MCP can be really be effectively utilized with AWS tech right now is the Q developer CLI command line interface. What I've noticed is it's offered broad support for model context protocol MCP servers right and we can get into the concept of you know versus a server. I know it's like a client server architecture. Um would love to hear a bit more about that too. Really what's really cool is it's able to offload a lot of the tasks that are very timeconuming but not necessarily linked to my direct day-to-day role. Right? Like as architects, we spend a ton of time in diagramming tools constantly, right? But our main focus should be on the actual architecture that we're really just translating into a visual format, right? However, when you're using these tools, a lot of it can be, you know, very granular aesthetic details. Getting an arrow to connect to another box, it can kind of drive you crazy, especially under a time crunch. I think we've all been there. >> Yeah. Well, with the QCLI, I've been able to have an MCP server dedicated just to generating architecture diagrams based on the current state of an AWS account. You can link it up directly in the QCLI and I've been able to get it to create architecture diagrams on the fly in PNG format, XML format, really any format of your choosing. Um, and it's been a game changer for my productivity. >> Yeah, that's amazing. Yeah, I I think in general if you the this is a a tool that's like hyper accessible to developers because virtually every uh ID including uh most of the Q tools uh Kira who folks uh may have heard about more recently uh come ship with some MCP native hooks to be able to plug right into servers. And so I think that's the adoption curve that we're actually seeing right now is like today it's primarily a tool used by developers and hackers to build out their own workflows. However, tomorrow and I think uh the the the future kind of what we we see coming together now is like actually using this to to to build a an enterprise agent. But that's very cool. I I I have the same experience with my coding companion. Um uh for me like some of the MCPs that are really valuable in my day-to-day work are diagram generation MCPs. Uh PowerPoint generation MCPS have been incredible for me. And then things that integrate with my uh like calendar and and mail and things like that uh go a long way for me since I'm spending most of my time communicating with customers. >> You're blowing my mind here. We could use MCP to create slide decks. >> Yeah. Oh yeah, for sure. >> If if spending time in diagramming tools takes up that big of a chunk of time, you can imagine how much time is spent building out slide decks and just preparing the aesthetics of what the real root technology is, which is what I think where we really want to invest our time regardless. Right? So, I'm already getting ideas here like the gears are turning even as we're talking about it how we could use MCP. >> Yeah. I mean, while we're on the topic, I I'll I'll highlight my my other favorite use case as a solutions architect here at AWS is so so many times we need to go into calls kind of showing customers an example of a piece of technology. I think we we refer to it as like art of the possible, but or like like a demo or a prototype. and being able to just completely kind of outsource that and prompt it into existence with MCP, particularly uh MCP servers that can integrate with AWS native services, AWS documentation. I find myself uh before a call, I used to spend uh hours prepping and now it's just a few prompts of putting together a demo or a benchmark and usually I can come into some with something pretty wellformed in a customer conversation. >> Yeah. And you bring up a great point with like AWS MCP servers as well. So we maintain our own repository of official AWS MCP servers. Correct. >> Uh we absolutely do. In fact, I was just uh down the road in another New York office with our PACE team. It's a team that works really closely with our our customers and essays in the field and they maintain a repo of 50 plus different MCP servers now including MCP servers to interface with AWS documentation and now to interface directly with AWS service APIs. Right? So, if you're an AWS customer or even if you're not an AWS expert and you need to be able to interface with AWS a little bit in your day-to-day job, it can be a real uplift uh for what you're capable of. I'll give you a specific example actually is at a I have a customer who um they have a central AWS team and they have developers kind of deploying into their Kubernetes clusters and these developers aren't AWS experts but they still need to do things like request IM roles and AWS resources and so being able to use MCPs that can programmatically read AWS documentation and then generate the Terraform code uh in a way that's like lease privileges for uh for a given IM role. Um, those are the sort of things where it's like that would have maybe been 10 Slack exchanges back and forth between a developer and a security engineer prior to this and now it can it can it's purely something that uh somebody in a in a dev team can kind of offload to the AI. >> Fantastic. So I want to hear a little bit more about some of the use cases that you've come across for MCP as well Trevor. >> Yeah, sure. I mean there there there are both sides of the coin. Use cases with customers and then use cases like uh that I use in my day-to-day like what what do you want to dig into first? I would say like is there any I'd say surprising use case for MCP that you didn't realize it could do and then you're like wait there's a server for this. >> Yeah I I do think that that happens a lot. In fact like MCP is one of the fastest growing protocols on Earth right now and there's uh tens of thousands of MCP servers being created every month. So this happens all the time where I'm kind of delighted by a new MCP server. I'll tell you one personally. Um, I'm a big user of there's an AWS um, customer partner called Perplexity and they have this deep research tool. And I never thought about what it could mean to uh, combine any deep research tool with a model because a and if folks who aren't familiar, a deep research tool is effectively a tool that just does a bunch of research on your behalf about a topic. But if you you pre-prompt a model with deep research, you're able to really automate a lot of the the configuration of a context uh window. So like for example, I might I might have a uh a deep research tool research a particular service or technology or or account prior to doing a build about it. So, I think it's it's those little nuances of really being able to inject more energy into a um into a model context. And I would say on the on the flip side with my customers, like I work in financial services and everybody wants to figure out, you know, every financial services customer has some huge data lake that they've been building out for years and years and they oftentimes have tools that investors or developers will use to interface with that data lake. And so virtually everybody wants an MCP server to frontend uh that that massive data lake that they that they've acred over the last decade. So that way they can interface with the data programmatically and just inject it into more people's workflows. And so I'd say that's a that's a really simple one where it could just be a single MCP server potentially fronting a really important data source and kind of unlocking uh what's inside of a data lake to folks that are not just developers who can write SQL code. >> For sure. And I I think what's really great too to tie it from like you know this is a technical solution that can solve business problems right and I think some of the obvious benefits that occur to me it's like amount of time saved right for developers anything else that comes to mind >> and speed too I I think there are instances where with AI you're able to uh consume and and aggregate information so much faster in in real time from a 100 sources at one time. You know it it's it's pretty easy to map it to just a human productivity workflow. Like if I'm a let's say my I'm an analyst and I'm researching five different companies in my portfolio. Think about all the different clicks that an analyst might go through of researching online looking at reports. There's so many different threads of that that now an AI can do all of those things at the same time. So what might have taken you an afternoon or day or a week can uh really be summarized in uh minutes maybe maybe an hour. Uh, and I think that I think that can't be overstated how how how that allows us to reinvest time in higher level tasks. >> Absolutely. And I want to like specifically highlight too like the AWS API MCP server as well, right? So now you're able to have a system that can basically run API calls on your AWS account as well. >> Absolutely. We it's been an awesome partnership with the AWS product teams that are kind of building out the interfaces for this. And recently we there's been a change here where the services were developing individual MCP servers. So there was like a Lambda MCP and a S3 MCP, but now they have just an AWS MCP that has all of our major services loaded into it. And so imagine that having one component that now unlocks your ability to paint with the canvas that is AWS primitives and services through natural language. I find it pretty uh pretty interesting how they basically did what MCP did >> for MCP it own way by consolidating that and standardizing across a variety of different services. >> Yeah. >> So dialing down into some of the more technical elements of MCP, right? I'd love to talk um you know at at least a high level, >> right, about some of the technical terminology that's used here. I know it's you know generally uh what would really an MCP server look like from a developer's point of view? Is this something that they build? How do they access them? >> Yeah. So there's there's a few ways to think of it. So first of all, the MCP protocol is a client server protocol, which if you've been in technology for a while, it's a very standard kind of architecture of a protocol where you have a client that initiates requests and then you have a server that understands those requests and and speaks back and forth. Now, an MCP server itself is re is generally just depending on its complexity, a few lines of Python code, right? And it's usually a a written in Python because that's kind of the language of AI. And inside of that server you define different things like for example you define tools. Tools might be APIs or systems that you uh read and write to. You define resources which often times are like static um instructions or or documents but those live inside the server. And so think of that server as a small almost like a a database that just understands all of the systems on the back end that it's automating. And in the example of AWS, that server would hold all these different API calls for all of our different storage services, for example. Uh so that's the server and then on the on the front end, you have a client and an MCP client can take many different forms. You could build your own. I'd say most customers are using a standard orchestrator like you said, a a QCLI is an incredible MCP client. Um there's some great open source ones out there as well. And that client it it it goes through a pretty typical handshake of just asking the server, hey server, what tools do you have? You know, and the so you might say, "Hey server, uh, what AWS services do you know about?" And it would spit back, I know about S3, I know about Dynamo DB. And all this happens transparently to the user. I think that's actually the magic of it is uh I we can talk a lot about MCP, but if there's one thing I could compel people to do that are listening, it is to like install a tool that's capable of using MCP and just use one server because it even though there's this complex technical handshake that's happening very much like a TCP handshake on the back end from the user experience level it really looks just like a uh AI is doing magic on your behalf like orchestrating and automating all of these things. I can really attest to that too, right? Like when you're working with Aentic AI systems and from my personal experience working with Aentic AI, especially with the advent of MCP servers now that are able to really level that playing field and standardize how we interact with these tools, it does feel like magic. No other way to describe it. >> Yeah. >> So, I want to talk about the learning curve for MCP as well, right? You're very well verssed on this topic, but how long did it take to pick up? I think what's great about AI these days is it's it's pretty easy to pick up most new concepts. You know, uh AI is really good at understanding technical things and explaining it to people in a very simple way. So, I would say really the the requirement to to in fact I back up. There's really two ways to think about MCP from a learning perspective. You you could think of just being an MCP user in which case you might be an individual developer that wants to build out your own workflows in your IDE or in a in a CLI tool. In that case, the learning curve is very very small. Like if if you're a developer that already knows how to use git, it's a pretty logical next step to learn how to uh configure and deploy an MCP server and use that in your workflow. So I say that that side of it of just deploying and using it is folks could do it in 10 minutes. You you could you could probably vibe your code your way into setting up an MCP environment. Now, if you actually want to start building MCP servers, that's maybe another layer deeper. However, like I mentioned earlier, there's like ten tens of thousands of MCP servers being developed every day, which means that there's millions of lines of referenceable MCP code already out on the internet. And so I I I I'm actually uh it's it's funny to say, but I've probably built I don't know 30 different MCP servers at this point, but I've never actually coded an MCP server. Like MCP builds all of my MCP, right? So I would say if if I'm going to recap that and give it somebody a phased approach, it's like step one is just get a tool that can use MCP and start playing around with the tool. And then eventually you might have a creative thought of like, oh man, it would be really interesting if I could also integrate this workflow into my calendar or this app that I really use or this notes tool. And then that's where you might build your first MCP server yourself to see what that looks like. But I think it's uh I would say it's it's something that anybody that has some very basic uh understanding of coding can get started with any time. >> And are there any specific learning resources that you found particularly helpful picking it up? >> I would really I personally referenced our official AWS MCP repo. So if you just search AWS MCP servers, it should show up on the top of your preferred search engine because within that repo, there's 50 plus MCP servers that I you can look inside of and see the code. It's very simple, elegant code that's relatively easy to understand semantically. So I think finding a a finding a server that you want to use that's open source is what I would suggest because then you can use it. You start by using it and then after you use it, you can look inside and see all of what's what's happening inside of the server to kind of deepen your understanding. >> Yeah. And you mentioned just like I feel users are spoiled for choice when it comes to the number of different options that they have when they're trying to pick an MCP server, right? Uh, do you have any recommendations on how to really narrow it down to find the right MCP server for your use case? >> I mean, I I I I can only say I I am a well, the right one for your use case should come pretty naturally, but there are some basic MCP servers I think everybody should be running, right? So, for example, you probably need a GitHub MCP server. You probably need some tool use or uh uh some sort of computer use MCP server that can open up a browser, right? uh you you almost certainly and I would say if you're just looking for a starting point, I'm a big fan of the AWS MCP servers. If you're listening to this podcast, maybe you're an AWS user and so that would be a really logical starting point is using some free tier services via via MCP on AWS. >> Makes a ton of sense. That's I think a pretty great starting point there. >> Yeah. >> Um do you have any general advice for folks who want to get started with MCP but aren't sure quite where to begin? >> I think everybody knows themselves well. For me, I'm the type of learner who I want to get hands-on where I want to watch I I I love watching videos and seeing people use the tool. And so, I would I would I would suggest that for folks, but again, I think the the number one piece of advice I could give anyone here is just it's actually a sneaky simple to use MCP. And uh you can talk about the theory of it and you can kind of try to understand it theoretically, but just get an IDE that's capable of it. deploy a single MCP server like an AWS MCP server and start using it. I think I think it creates a flywheel because seeing it work is is much makes it much more real than trying to understand all of the complex intricacies of uh prompts being converted to API calls back and forth. >> Of course. And I think another good question too that really I think that kind of sparks up for me is what exactly would you say because I know with MCP, right, it on paper really does closely resemble like an API in a way, right? So what's the difference between using MCP compared to just like a standard REST API? >> Yeah, I I it's a it's a great question. In fact, this is one of the most common questions when I'm running an MCP workshop that a customer will ask. >> Uh there there's a lot of similarities. I I I think that there's there's two big things that I think about when it's when I'm thinking MCP versus API. First of all, uh not everything has an API. And so there are some systems that you may want to interface with that would require a different type of orchestration. So imagine you're a company. I'll give you I'll go back to the example of I'm a financial services company that has like a data API, right? Uh imagine now it's late 2025 2026. you've got a few agents in in deployment and um without something like MCP what that looks like is you know those let's say have 10 agents deployed each of those 10 agents will have their own integration so you have 10 integrations built with this data API and that is that may not sound that bad at surface level but when you think about enterprise tooling uh enterprises really care about like does the right person with the right permissions have access to this data and there's also a productivity angle of like do you really want to be building the same API integration 10 different times with uh what what what was the standard before MCP which was uh tool calls within LLMs. Fast forward to MCP. Now you have the capability to build a single MCP server for a data source in this example and have all of your agents plug into that in a system systematic way. And so I I think it it it's it it prevents the kind of bespoke nature of tool calls within uh agents and it also just makes it such that it's easier to orchestrate agentic behavior because uh uh the other pattern here is prior to MCP to build something agentic you needed an agentic framework. So you would have needed a strands or like a lang chain or there's many different agentic frameworks out there but that's kind those are pretty complex. Um, and even still that that there there's some development overhead involved with those tools via there's just much less code to build an agent when you're using MCP. It's it's a very natural language driven and uh I think that that allows people to move faster and I've even seen instances where people who are non-technical um like we have a tech an wonderful technical program manager on our pace team named Anita who uh wasn't a programmer in a past life but is able to produce all of these incredible demos and assets using MCP because it's not so um technical and difficult to wrangle for a non-developer. >> Yeah. Really democratizes it. >> Yeah. even for folks who maybe don't come from a software development background and I think that's one of the huge advantages um of Ajetic AI in general, right? Is how it really opens the doors for people too. So very exciting stuff. I do want to ask a little bit about you too. How did you initially get into tech? >> For me from a a young age in high school, I I kind of saw it as a way to start a career and and get uh get my feet underneath me from a from a life perspective. So I I my path was I joined the military whenever I was 18 years old and I joined the US Air Force as a effectively a network administrator. So I started running cables through brick buildings with uh that you couldn't bring cell phones inside of they called skiffs and uh configuring routers, switches, firewalls for the US military. And then it's kind of just been a natural progression from there. I've always been a very ever since I got into technology I I I bet a lot of folks listening would empathize with this. I just kind of wanted to understand as soon as I understood TCP IP and packets and networking, I just wanted to understand how everything fit together. It's always it's always been like a big puzzle to me of like who wa the internet works. Whoa, like people are deploying all these different applications. And I think it's been the pursuit of just trying to understand how all this stuff works has led me to transition from initially networking then into the data center for a few years and now I'm very lucky to have gone through the cloud uh era and be right in the in the middle of it all during this AI uh this AI thing that we're going through. >> What a time to be in the business, right? >> It's it's what I I feel grateful lucky every day. I I I always thought like, man, like I I got out of college and and kind of started my career after the the early 2000s and like the the.com boom and I always thought like, oh wow, wouldn't it have been really cool to be a part of some big transformation like that and here we are in the I think the the first inning of what is going to be a probably decadesl long transformation and we get to be part of taking it to customers. I mean like what what else could you want as a technologist? It's incredible. And I'm curious, too, how that early career experience also shaped what you do today. Were there any major takeaways from those early days of putting together chicken quesadillas? I read the LinkedIn. I did. >> Oh, yeah. That's right. I did. I worked at Taco Bell briefly. That had no connection to uh getting into tech, but it was a it was a fun job. And I do love a I do love a good quesadilla. I I think there was an epiphany that I had a few years into my career right as I got out of the military which was which was effectively that technology like a lot of network engineers I fell in love with the tech I fell in love with you I was thinking of packets in my dreams you know and uh I was deep in wire shark and I just I was loving technology for the sake of learning and that's good it's incredible to learn but I had this moment towards the end of my air force enlistment where I it hit me that like all of this technology is just a means to an end. It it is like the these are just uh these are insignificant compared to what they can accomplish. And that may seem like a a silly realization, but for me, what that led me to was understanding that I need to constantly be refreshing and throwing out and learning new things because all these technical skills that I have are actually in service of the US government or a financial services customer or Amazon's business and really the technology is just a means to an end. I think that framing allowed me to let go of that identity of being a network engineer and open up to the idea of just being like a generalist architect that has been able to evolve with the industry around me. >> What do you think's next? I know MCP is one of the biggest topics at the forefront of everyone's mind right now, especially with the advent of Agentic AI, but what do you think is on the horizon? In my mind, it's actually pretty clear what's on the horizon because as you use AI tools, you can you can still feel some of the sharp edges. Like sometimes the AI won't do exactly what you want it to do. And or or sometimes even in in the case of MCP, this will happen a lot actually where you ask an AI to build an app for you and instead of building exactly what you asked, it will build like a mock version of what you asked and it will work. The demo will work, but it won't be like the working functionality. And so I think those are those are just some small gaps not in MCP but actually in the intelligence workhorse of the app in in the model itself that naturally over time we should overcome. So I I believe by this time next year, just seeing how far a agentic coding models have gone in the last 12 months, we'll be at a point where you you truly can in a hands-off way um have AI do a lot of of coding and um kind of heavy lifting of writing application code for you. And we're there today to an extent, but I I do think that we're going to take our hands more and more off the wheel as developers as we build more trust and the models become more capable. Um, outside of that, it's kind of it's kind of hard for me to to imagine beyond that because, uh, in this world, thinking more than a couple years out feels pretty impossible cuz I just know how different it is now than it was 12 months ago. >> For sure. And I don't think either of us know the future. But I think what we can say with certainty is that the present is pretty incredible. To give you some context, on the train ride up here, I was able to build an entire commute planning application in Python using Curo, and it took me about 30 minutes, not even. I'd say that's a very generous estimate. >> Yeah. >> Um, it's incredible how quickly you can get from just an initial idea to a very very fleshed out prototype with the power of AI tooling like Hero. It's it's pretty amazing. And that's impossible without MCP under the under the covers. >> MCP under the hood. >> Yeah. It's got a it's got a what would you say? For me, it feels like there were there were coding assistants last year and now there's these MCP enabled coding agents. I would say my coding productivity feels like it's probably four or 5xed in this paradigm where it's we've evolved from autocomplete of code to actually agentic coding. I do feel as a force multiplier. >> I think it's been an absolute force multiplier, right? And I think to put it kind of more in layman's terms, what I've noticed is just how much less time I have to spend doing a lot of I think some of that more menial development, right? So like writing unit tests, writing documentation, right? Like you can very quickly get a readme file that's able to talk about like your database schema prerequisites, any type of dependencies that you need to have installed. I think that's a big thing too if you're working on a new project that requires a couple new dependencies, right? like just fighting with the terminal in your IDE to get to a stage where you have everything installed so the application can even run, right? Having just basically like a battle buddy on the sidelines here that can help talk through what you need to do in order to get all those dependencies installed and running. You know, even just some of that basic foundational stuff. Um, so imagine how easy it is now if you're a junior developer or just someone getting into the into the craft, right? How much easier it is to get started with the power of AI. >> Absolutely. I I I would say my I I would have called myself a novice coder until until AI came out a few years ago. To be honest, I I was writing some simple Python script and now it's like I'm constantly amazed at what what what you're capable of. And I I I I think I think it will drive a lot of people to be just to maybe follow in our paths and become more generalists actually because I I think I think you'll the the role of a developer might transform a little bit more into being focused on converting requirements into the the instructions that deliver the code rather than just taking the instructions from a a project manager or a PM to implement it yourself. >> I think that's the trend that we're really going to be seeing as well, right, is everyone can be a generalist with the power of AI. I think it really does help level that playing field. I'm very excited to see what the future holds. Um, this is some really, really impressive technology. So, Trevor, thank you so much for coming on the show. Um, I've certainly learned a lot. It's been fantastic hearing your insight on MCP. Um, I hope to have you on again sometime, but any parting thoughts for the crowd before we uh before we part ways tonight? >> No, I I'm going to sound like a broken record. No parting thoughts. Thanks for having me on. I would say if you haven't done it yet, folks, download an MCP server uh in your ID. give it a try and and reach out to Chris and I. Let us know uh how it went for you. We'd love to we'd love to hear what your experience has been like. >> Absolutely. Thanks everybody for watching. [Music]