What an MCP Server Actually Does

An MCP server gives an LLM a standard way to reach your data and run tools against it. Every API is built differently, and the Model Context Protocol puts one interface over all of them, which is why Anthropic's standard gets compared to USB C.

This episode covers what the standard defines, how tools and resources set the boundaries on what an AI can do with your systems, and what you get by writing a server yourself. We built one for our own AI-ready data model and hit a context problem along the way: the model would skip the business context files and bluff its way through an answer, then admit it never read them when asked directly. The fix was a token buried in those documents that the server checks before it will run any other tool.

If you work in data, BI, or analytics engineering and you are weighing a vendor's MCP server against building your own, start here.

Like and subscribe for straight talk on data, AI, and running a modern business on real numbers.

Blue Margin builds and manages data platforms on Microsoft Fabric and Power BI for mid-market and private equity backed companies. Our Managed Data Service consolidates your systems into one governed source, delivers the dashboards your team runs on, and maintains them as your business changes. Learn more at bluemargin.com.

Transcript

SPEAKER_02 0:06

Welcome to the Dashboard Effect Podcast. I'm Brick Thompson.

SPEAKER_00 0:09

And I'm Landon Oaks.

SPEAKER_02 0:10

Hey Landon. So today we wanted to talk about MCP servers. I'm sure most of our listeners are hearing about those. They're a big part of how the uh how people using LLMs are able to um connect to data and pull data into the LLM to be able to get answers to questions that way. But I wanted to talk just a little bit about at a high level, what is the structure of an MCP server? What's the point of using an MCP server as opposed to just connecting to an API?

SPEAKER_00 0:39

Yeah, it's a good question. It's um I mean in a nutshell without going too deep here, it's essentially a way for AI to actually interact with your data or even run certain tools, do things kind of autonomously. Um but it's very universal, right? It's standardized. Exactly. Yeah. So it's universal across all AI can kind of work with it. Um you get a lot more control over it. You know, if you think about APIs in general, um, you know, they're all different, right? Every single one. So without that MCP server kind of being a translation layer, you're gonna have to write custom for every single API that you wanted to interact with. Um and so now a lot of these like bigger companies are coming out with their own MCP servers, they're making it publicly available. I'm sure a lot of people use them already. Um so it's just a really cool way to get AI to do more for you, essentially.

SPEAKER_02 1:35

Yeah. Well, and it's a standard that Anthropic came up with, I think, almost two years ago now. Yeah. Maybe not quite. Year and a half, something like that. Maybe maybe it's two. It's about two. Um that that defines how you should build an MCP server so that your LLM will be able to use it as a tool. So it defines what it can get, how how it uh how it can interact with the things that uh that you know, the data and so on that it's getting from these different systems. Um so there's tools, there's I think they call it resources, which would be the data, um and uh sort of rules about how to interact. Sort of like you know, the the analogy that everybody uses is like USB-C. Like everybody knows how USB-C works. If you build it right, then you can just plug it in and use it with whatever peripheral you're using.

SPEAKER_00 2:24

Exactly.

SPEAKER_02 2:25

So um also people build custom MCP servers a lot. And we we do here, we've talked in past episodes about our platinum layer, which if you're in the data world and you know what the medallion architecture is, we've talked about that at length where you take raw data and then you manipulate it, modify it, and eventually get it from bronze to silver to gold layers in your medallion architecture, gold being the layer that allows you to do good BI and reporting. And then we're extending that to what we're calling a platinum layer, which organizes the data in a way that uh an LLM can interact with it well. Uh but in order to give the LLM access to that platinum layer, we've had to write an MCP server. And um, we've done some things in there, you know, following the standard, but also um extending a little bit for our specific um need. Maybe you can talk about a couple of those.

SPEAKER_00 3:21

Yeah. Yeah. And I mean it's it's really cool being able to write your own because you're able to customize it, right? You know, almost the sky's the limit at that point. Um so like a couple of cool things that we ended up doing that's been helping a lot is we will basically force it to read context about the business and the data that it's that it's getting so it actually knows what it's looking at rather than just kind of being thrown data and guessing, trying to infer what it thinks it is.

SPEAKER_02 3:48

So and and just to we've talked about this in the past, but just to remind people what what we're talking about there, so we've created a data model that LLMs can use very easily. It's very friendly for LLMs. But in addition, we add a bunch of context files to the data lake house that describe the business and what types of questions people might be asking, and depending on who the audience is, you know, what they might actually be looking for, um, those types of things. And we embed that into the lake house. But then we need to be able to have the MCP server force that context back to the LLM so it can give better answers.

SPEAKER_00 4:25

Yeah. Yeah. And that was that was one of the things. I think we've talked about it before, so I won't go too deep. But um when we were first just asking it to read that, it, you know, we had times where it was just wouldn't. It it kind of bluffed its way through it. Once you asked it directly, it said, no, I did not read that. Well, I'll go read it now. Um so one of the ways we got around that was essentially just giving it like a little token that it finds in those documents. And if it doesn't have that little, like almost password, it can't call any of its other tools. And it'll just get back, hey, you need to go read this. You did not get the token.

SPEAKER_02 4:56

That forces it into the LLM's context.

SPEAKER_00 4:59

Exactly. Yeah.

SPEAKER_02 4:59

And when we did that, we found we got much better results. As you said, I I think it was Claude actually that gave gave a bad answer, and our engineer who was testing it said, Did you read the context files? Oh no, I'm sorry.

SPEAKER_00 5:11

Yeah, exactly.

SPEAKER_02 5:12

Don't you remember the instruction that you have to read the Yeah, yeah, I just didn't.

SPEAKER_00 5:16

Yeah, yeah. Trying to take a shortcut, right? Trying to get faster at it. Yeah, yeah. So yeah, on top of that, too, you know, there's there's other things as well, like we can build in security. You know, you can tell it basically the tools that you create is what the AI can do. So you're able to get pretty good control over it, not have to worry, like, oh, is this AI gonna go drop my database or do something fishy there? Right. Because we didn't give it a tool to do that, so it can't, right? So um you get a little bit more control in in terms of what it can and can't do.

SPEAKER_02 5:48

Yeah, yeah. And and we can build stuff in there like here's here's how we want you to query the platinum layer, those types of things. So it it makes it uh much better. So the MCP server is key. I mean, people are using it if they're if they're using connectors in uh ChatGPT or Cloud, um, probably other systems too, other hardnesses, they're running across that. Um but I think it's important to know that you don't have to be limited to just something that a software uh a SaaS system has developed for you. You can actually develop these yourself. Um you may not be able to develop it on someone else's system, but if you've got access to an API, you can definitely develop an MCP server that does what you need it to do.

SPEAKER_00 6:30

Exactly. Yeah.

SPEAKER_01 6:31

All right. Anything else you wanted to cover there?

SPEAKER_00 6:34

I don't think so.

SPEAKER_01 6:34

All right. I think we'll wrap this one up then. Thanks.

SPEAKER_00 6:37

Great, thank you.

unknown 6:38

All right.