Latest / Microsoft Cloud IT Pro Podcast / Episode 407 – Model Context Protocol (MCP) Part 1: The USB Port of AI Communications
Transcript
- 0:03Welcome to episode 407
- 0:05of the Microsoft Cloud IT Pro podcast recorded
- 0:08live on 07/25/2025.
- 0:11This is a show about Microsoft three sixty
- 0:14five and Azure from the perspective of IT
- 0:16pros and end users, where we discuss the
- 0:18topic for recent news and how it relates
- 0:20to you. In this episode, we dive deep
- 0:23into the USB port of AI communications
- 0:26as we discuss model context protocol or MCP.
- 0:30This is a game changing specification that's extending
- 0:34the capabilities of large language models or LLMs
- 0:37to help you bring your data together into
- 0:40a central LLM of your choice. We'll talk
- 0:43about what they are, where you can use
- 0:45them, and things you may want to think
- 0:46about when it comes to using these MCPs.
- 0:49So let's dive into the show.
- 0:54Scott, we're between vacations. I'll pretend I'm surprised.
- 0:57Yeah. I know. You just got back from
- 0:59vacation. I'm about ready to leave for vacation.
- 1:02Summer recording is always interesting. I had a
- 1:04good vacation.
- 1:06We went to
- 1:08we we did Oregon,
- 1:09Northern California
- 1:11and just did a a ton of driving.
- 1:13So we did about 2,000 miles in ten
- 1:15days. How many kilometers
- 1:17for our non American friends?
- 1:19What is it? It's multi multiply
- 1:22by, what, three? Something like that. I don't
- 1:24know. I don't know. It's a bunch of
- 1:26kilometers. But yeah, good trip, lots of hiking.
- 1:28We went and saw like the redwoods and
- 1:31did,
- 1:31Crater Lake and all sorts of stuff. 3,218
- 1:36kilometers is what you drove. Not too shabby.
- 1:39Thank goodness for rental cars and somebody else's
- 1:41mileage. That sounds good. I'm kinda I'm jealous.
- 1:43Yeah. I have never been out, like, Northern
- 1:45California,
- 1:46Oregon area. Have you ever seen The Goonies?
- 1:50I don't you're gonna make fun of me
- 1:51again. I don't know that I have. Then.
- 1:53Another one of those. If it's been a
- 1:54while if I have, it's been a while.
- 1:56I'll make you set up the projector outside.
- 1:58I'll come over. You can cook me dinner,
- 1:59and we'll watch you at least. The Goonies?
- 2:01So that was kind of our one of
- 2:02our last stops on that trip was staying
- 2:04up in Cannon Beach where Haystack Rock is,
- 2:07and it's like the big rock that the
- 2:08pirate ship floats out of at the end
- 2:10of The Goonies and things like that, so
- 2:11that was kind of fun too. The other
- 2:13really nice thing
- 2:15about
- 2:16that part of California, like the very tip
- 2:19Northern California up through Oregon, you're in the
- 2:21Pacific Northwest in the summer,
- 2:23it is not like Florida in the summer.
- 2:26So Florida in the summer, like I'm looking
- 2:28right now, it's
- 2:3088% outside with 82%
- 2:32humidity and feels like 106%.
- 2:34It is miserable and you do not want
- 2:36to be outside.
- 2:38In that part of The United States, it
- 2:41is
- 2:42because you're right on the coast, so you
- 2:43get that coastal fog in the morning. Some
- 2:45days you can't even see the sunrise or
- 2:47the sunset because the fog is so thick.
- 2:5070
- 2:51degrees, 60 degrees,
- 2:53it all just hits different. It was absolutely
- 2:56beautiful. I got super sad when I got
- 2:58there because I forgot to pack a pair
- 2:59of shorts and I was like, Oh my
- 3:00gosh, it's going to be like seventy, eighty
- 3:02degrees, it's going to be hot. It was
- 3:04perfectly comfortable every day in a sweatshirt
- 3:07and long pants and all that stuff.
- 3:10So it was really great, good little change
- 3:12of pace. I'm ready to go back, I
- 3:13think, on vacation. Doctor. On vacation? See, I'm
- 3:16looking forward to it too. We're going up
- 3:18north to Michigan, and then we're going out
- 3:20to Denver. And same thing, I'm like, it's
- 3:22not Florida in the summer.
- 3:24It'll be beautiful. With that, can you use
- 3:27an MCP
- 3:28something or other to help you plan your
- 3:30vacation? To plan my vacation? Was that was
- 3:33quite a stretch there, Scott, for it. For
- 3:35a segue, you were looking for one? Yeah.
- 3:37It all went sideways. You mean a Microsoft
- 3:39certified professional? Do Microsoft certified professionals? We like
- 3:42acronyms. Right? It's what an MCP used to
- 3:44be. I guess it still is. I was
- 3:47thinking more like model context protocol
- 3:49and
- 3:50some of the new specs and things have
- 3:52been coming out of anthropic and then the
- 3:55really cool nifty functionality
- 3:57that's been coming out along the way. Yeah.
- 3:59I mean, I suppose you could. Right? Like,
- 4:01with MCPs
- 4:03and kind of the goal, like you said,
- 4:04model context protocol, is that this kind of
- 4:07extends LLMs to other data sources. So I
- 4:09suppose
- 4:10maybe if there's MCPs
- 4:12out there to connect to
- 4:14some type of data that you might want
- 4:15to use to plan your vacation or something
- 4:17that you're planning your vacation from, you probably
- 4:19could use an MCP to plan your vacation,
- 4:21Scott. Probably. What is an MCP? If it
- 4:24hooked up to the right set of data
- 4:25sources. Absolutely. Yeah. With vacations, this is gonna
- 4:28be maybe the should we give people context?
- 4:31We are doing a longer episode that we're
- 4:33gonna break into two parts today because of
- 4:35vacations. So we're gonna kinda start off with
- 4:37what is an MCP
- 4:38followed by how do you use MCPs? We're
- 4:41jumping ahead of ourselves.
- 4:43So what is an MCP, Scott? What is
- 4:46the model context protocol
- 4:48from Anthropic,
- 4:50creators of also Claude? Model context protocol is
- 4:55a specification,
- 4:57like first of all, right? It's a Doctor.
- 4:58Joshua Klooz (zero 50 three:fifty three): Yep. Doctor.
- 4:58Joshua Klooz (zero fifty three:fifty three): Framework
- 5:00and architecture
- 5:01for
- 5:02developing
- 5:03servers
- 5:04and clients
- 5:06that can interact with large language models
- 5:10and LLMs and become part of
- 5:13agentic workflows
- 5:15for LLMs. So they can surface very specific
- 5:18pieces of functionality.
- 5:20So for example,
- 5:21you might have OpenAI out there and you're
- 5:23using ChatGPT
- 5:25four o, and you're chatting with that, and
- 5:28that thing was trained on a specific set
- 5:31of data.
- 5:32It could be that you're maybe grounding it
- 5:35in some additional data. Maybe you're doing rack
- 5:37or things like that. But maybe you need,
- 5:40like, additional functionality
- 5:42or more
- 5:43system or service
- 5:45specific functionality
- 5:46to do things. For example, let's do like
- 5:49a pie in the sky thing. So
- 5:51I have a Notion database, and one of
- 5:53my Notion databases is recipes, where I capture
- 5:57recipes that I find on YouTube. There's a
- 5:59sub database that actually has, like, the recipe
- 6:02itself. Here's the steps. Here's the ingredients,
- 6:05all those kinds of things along the way.
- 6:06So one of the things you might be
- 6:08able to do is go out and ask
- 6:10an LLM,
- 6:11hey. Can you give me a new recipe
- 6:14for
- 6:15roasted chicken, roast potatoes, asparagus,
- 6:18and make it a little spicy? And what
- 6:20would be good if it had, like, these
- 6:22flavor properties?
- 6:23Well, if you had it tied into a
- 6:25tool with an MCP agent, you could also
- 6:27extend that prompt and you could say, once
- 6:30you've have that recipe and I've iterated on
- 6:32it, great. Hey, that recipe looks good. Can
- 6:34you take it and save it in my
- 6:36Notion database over here? And then because now
- 6:39you have the ability to
- 6:41have that Notion specific
- 6:43functionality
- 6:44tied into
- 6:45your client that you're chatting with that LLM,
- 6:48now you can start to chain that together
- 6:50and you can, like, push that information down
- 6:51from the LLM
- 6:53out to an external tool. Or you can
- 6:55do this in other ways. Like, let's say,
- 6:57for example,
- 6:58in Azure land, there's an Azure MCP server,
- 7:02and that server lets you do things like
- 7:05list the resources
- 7:07in my subscription.
- 7:08Great. Like, list all the resources.
- 7:10Now, once you've listed the resources, go find
- 7:12the properties for those resources. Now, once you've
- 7:14found those resources with those properties, maybe take
- 7:17them and write them out to this other
- 7:19thing. Like, send them to Slack, send them
- 7:21to Notion, send them to OneNote, something like
- 7:23that, so that you can continue to have
- 7:24it go further. So you end up in
- 7:26this world where you can extend LLMs
- 7:29with
- 7:30domain specific knowledge and domain specific functionality that's
- 7:34built by developers and by these services, and
- 7:36then you can chain it all together because
- 7:39now the clients that are interacting with these
- 7:41LLMs and have the ability to integrate with
- 7:43MCP servers,
- 7:45they can also know, like, oh, hey. I'm
- 7:47looking at an agent or a tool that
- 7:50surfaces this piece of functionality.
- 7:52Therefore, go talk to this, chain it into
- 7:53this, things like that. So that's all really
- 7:55cool because you can do that just through
- 7:57a singular client
- 7:59and kind of like a real time, like,
- 8:02back and forth kind of interaction.
- 8:04The other way that these things can be
- 8:06used is as part of agentic workflows.
- 8:08So rather than me being like, hey. I'm
- 8:11gonna sit here and prompt you in real
- 8:12time, maybe I just have a prompt that
- 8:14kicks off overnight and does something in the
- 8:16background for me. And it's like one agent
- 8:18talks to another agent, talks to another agent,
- 8:20talks to another agent, talks to another agent.
- 8:21So if anybody's maybe been playing around with
- 8:24some of the agents or tools, like, they're
- 8:26called different things all over the place. Like,
- 8:28in in Microsoft three sixty five Copilot, they're
- 8:31called agents. If you go into like Copilot
- 8:33Studio, you've got these things called tools and
- 8:35other bits and pieces. Like this is a
- 8:37more rich interaction
- 8:39surface
- 8:40on top of those things that have been
- 8:42there before. I think one of the weird
- 8:44things about MCP in general is when I
- 8:46think about it, I think about more from
- 8:48like the end user side, like how do
- 8:49I as a user get started with it?
- 8:51Well, there's a whole bunch of LLM clients
- 8:54out there, like Copilot or heck, ChatGPT itself,
- 8:58that don't allow you to actually interact with
- 9:00MCPs,
- 9:01but others do. So if you have maybe
- 9:04like Claude or Perplexity on the desktop,
- 9:07then those let you run and integrate with
- 9:10MCP servers, both local servers and remote servers,
- 9:14which that's a distinction that we should talk
- 9:15through, tools like
- 9:17AI driven IDs like Cursor, WindServ,
- 9:21Versus Code as well, right, if you're just
- 9:23looking to, like, pick this up and and
- 9:25get going and get free free stuff along
- 9:27the way. So
- 9:28I think they are
- 9:30a
- 9:31weird new enabler, Like, they're definitely, like, cutting
- 9:33edge and we're on some, like, precipice or
- 9:35cliff, and it's not clear, like, does the
- 9:37ground continue to extend in front of us?
- 9:39Do we fall off the edge, and what
- 9:40does that look like? But it's definitely, like,
- 9:42exciting just to, like, be able to get
- 9:44hands on with some of this stuff and
- 9:46leverage it and find ways to
- 9:48use it in
- 9:50your day to day workflows, be it like,
- 9:52hey, you're a consumer. Like I said, maybe
- 9:53you're sitting here and you're doing, like maybe
- 9:55you're like a content creator and you're doing
- 9:57video creation and you say, okay, hey, ChatGPT,
- 10:00create my script. Well, great. Go into Versus
- 10:03Code, have a chat with ChatGPT four zero,
- 10:05create my script, now take my script and
- 10:06push it out to here, and let me
- 10:09leverage it that way or have it and
- 10:11iterate it on it, things like that.
- 10:13Or just in my day to day professional
- 10:15job, I'm always looking at
- 10:18documentation,
- 10:19code, samples, all those kinds of things. So
- 10:21there are very specific
- 10:23MCP
- 10:24server implementations that I can integrate with a
- 10:27client
- 10:28and be able to really kind of just,
- 10:30like, accelerate
- 10:32and augment those workflows within my day to
- 10:34day, which is very exciting. Right? Like, I
- 10:37think it's, like, I like cutting edge and
- 10:39raw, and it's always fun to get hands
- 10:41on with a new tool or a new
- 10:43toolset. A lot of these and where I've
- 10:45started playing with MCPs too is you mentioned
- 10:48like Claude or OpenAI or Copilot.
- 10:50It's you're limited in what data you can
- 10:53access. You can go build an agent in
- 10:54Copilot,
- 10:55but in that case, Microsoft has very
- 10:58defined connectors. They're starting to get more with
- 11:00some of the custom connectors and different things
- 11:03there, but like OpenAI. OpenAI,
- 11:06the data sources,
- 11:07kind of before MCPs, were a lot more
- 11:09limited. OpenAI is set to go look at
- 11:11the web or look at this data or
- 11:13use this model. Claude, same thing. This is
- 11:16really opening up that ability to say, now
- 11:19I want Claude to go look at all
- 11:22this other data that's more specific to me.
- 11:25Like you mentioned, I want Claude to go
- 11:27be able to look at my data in
- 11:29Notion, or I want Claude to be able
- 11:31to go look at the data in a
- 11:33SQL database. I want Claude to be able
- 11:35to go look at
- 11:37data over in
- 11:38Microsoft three sixty five in SharePoint land. It's
- 11:41allowing you to bring your own it's almost
- 11:44like bring your own data. Do we have
- 11:45a BYO instead of BYOD for devices? It's
- 11:48BYOD for data. Bring your own data into
- 11:51these different LLMs so that as you're querying,
- 11:53you can
- 11:54get different responses, better responses, more accurate responses,
- 11:58or responses very specific to
- 12:01the data that you want these LLMs to
- 12:03be able to parse over and
- 12:05pull these results from or these responses from?
- 12:08It's that domain specificity
- 12:10and the ability to plug and play domain
- 12:13specificity. So Anthropic, the folks who make Claw
- 12:17and SONNET and all that stuff, they are
- 12:20the ones who came up with this specification
- 12:22and pushed it out there into the world,
- 12:24and one of the ways that they frame
- 12:26it is
- 12:27MCP
- 12:28is the
- 12:30USB port of
- 12:33AI communications
- 12:35and kinda AI interaction.
- 12:37So if you think about that, like, if
- 12:38you think about, hey, I have this LLM
- 12:41that's sitting here, and now it's got an
- 12:43infinite number of USB ports on it. Just
- 12:46think about, like, all the things that you
- 12:47plug into your computer to help you do
- 12:49your job better. You plug in the dongle
- 12:51for the mouse. You plug in your monitor.
- 12:52You might plug in a dock. I've got
- 12:55an audio interface. I've got a camera. All
- 12:57these different things sitting in front of me.
- 12:59Well, those abstractions
- 13:00also apply in this world of MCP, where
- 13:03you've got an LM that's now kinda sitting
- 13:05there, and you've it's got, like I said,
- 13:07infinite USB ports, and you just start plugging
- 13:10in that domain specific knowledge
- 13:13and now the cool thing about the spec
- 13:15and the way it's laid out is it's
- 13:17kind of built around this core set
- 13:20of building blocks that they see and that
- 13:25sit out there. So clients can have a
- 13:27whole specific set of functionality
- 13:29where they can plug into a couple underlying
- 13:32building blocks within the MCP protocol itself. So
- 13:35you've got resources, prompts, tools,
- 13:39there's some other more esoteric things that I
- 13:41don't think I've seen involved in too many
- 13:44MCP implementations like sampling and roots and elicitation
- 13:48and things like that. But there's kind of
- 13:50like these core building blocks in
- 13:53tools,
- 13:53resources, and prompts. So tools are these things
- 13:57that
- 13:58are specifically for AI actions
- 14:01and AI interactions.
- 14:03They are controlled by the model itself, so
- 14:06the LLM is doing things. So that might
- 14:08be like I have an MCP server maybe
- 14:11that searches for flight information. Right? Maybe it
- 14:14integrates with, like, kayak.com
- 14:16or something like that. Well, that could be
- 14:18controlled by the model where I could have
- 14:21a tool in an MCP server that's directly
- 14:23integrated with the Kayak API where I could
- 14:25say, hey. Go search I just tell my
- 14:28LLM. Go search for flights on this date
- 14:31with this carrier from a to b. And
- 14:33based on the context of, oh, I see
- 14:35the user is searching for flights, and I
- 14:37see I have this MCP server over here
- 14:39that I can attach to, and it's offered
- 14:41me an agent with this capability. Let me
- 14:43go use that. Let me go get that
- 14:45domain specific knowledge. So that could be like,
- 14:47like I said, something corny. You're searching for
- 14:49flights. You're working on your scheduling, calendaring,
- 14:52simple things like that. Maybe you're using it
- 14:54as a way to send messages back and
- 14:57forth in Slack or things like that. So
- 14:58that's one of your building block is tools,
- 15:00which are model
- 15:05controlled. Do you feel overwhelmed by trying to
- 15:07manage your Office three sixty five environment? Are
- 15:09you facing unexpected issues that disrupt your company's
- 15:12productivity?
- 15:13Intelligink is here to help. Much like you
- 15:15take your car to to the mechanic that
- 15:17has specialized knowledge on how to best keep
- 15:19your car running, Intelligink helps you with your
- 15:22Microsoft cloud environment because that's their expertise.
- 15:25Intelligink keeps up with the latest updates in
- 15:27the Microsoft cloud to help keep your business
- 15:29running smoothly and ahead of the curve. Whether
- 15:31you are a small organization with just a
- 15:33few users up to an organization of several
- 15:36thousand employees,
- 15:37they want to partner with you to implement
- 15:39and administer your Microsoft Cloud technology.
- 15:43Visit them at inteliginc.com/podcast.
- 15:47That's intelligink.com/podcast
- 15:54for more information or to schedule a 30
- 15:56call to get started with them today. Remember,
- 16:05And then you have these things called resources,
- 16:08which are application controlled or controlled a little
- 16:11bit more by the
- 16:13server implementation, the MCP server,
- 16:15server implementation itself. So those are all for
- 16:18contextual data. So those would be things like
- 16:21documents,
- 16:22calendar events, could be emails,
- 16:25could be other domain specific data things like
- 16:27that like anthropic I think in one of
- 16:29their examples uses like weather data
- 16:31as a thing that's there in that first
- 16:33one where I had a tool I said
- 16:35go search for the flight and go do
- 16:37these things for me The other thing that
- 16:39I could do is then I could have
- 16:41a resource
- 16:42that goes and actually reads out of my
- 16:43calendar. So don't just go search for flights.
- 16:46Maybe I don't say search for flights between
- 16:47these dates where I say search for flights
- 16:49between these dates
- 16:50that don't conflict with my existing calendar. And
- 16:54if there's an integration with, like, Google Calendar
- 16:57or Outlook or things like that, then you've
- 16:58got all that. And then finally, a third
- 17:00building block that kinda comes into play with
- 17:02these that you'll see in a bunch of
- 17:03client implementations
- 17:05is prompts.
- 17:06So these are for, like, interactive
- 17:09templates basically.
- 17:11So it's all user controlled.
- 17:13So when you're kinda you can basically ground
- 17:17these things and give them a set of,
- 17:18like, base instructions
- 17:20that help them understand
- 17:22where you want them to go with it.
- 17:24So if I'm back to, like, that thing
- 17:26of kayak and search for flights, well, maybe
- 17:28I'm planning to go on a vacation or
- 17:29a business trip. So I could provide some
- 17:31very specific grounding that says, I'm planning a
- 17:35family vacation
- 17:36for a family of four. My wife likes
- 17:38to do this. My kids like to do
- 17:40this. And now I don't have to put
- 17:41all that into as I'm prompting the LMS
- 17:44as I go along. It's just like a
- 17:45base set of instructions that it knows. Maybe
- 17:48you've got, like, multiple
- 17:50MCP servers that you're talking to. Maybe you've
- 17:52got one for, like, Kayak. Maybe you got
- 17:53one for Google flights. Maybe you've got one
- 17:55for, like, a carrier like Delta. Maybe you've
- 17:57got one for, like, hotels like Marriott, things
- 17:59like that. And you can actually tell it
- 18:01in the instructions like, hey, I I really
- 18:03prefer you to book and look for hotels
- 18:05with Marriott or like me, like, I tend
- 18:07to fly with Delta a lot. Like, please
- 18:09only pick Delta flights with
- 18:11no more than two stops and that go
- 18:13through these hubs kind of thing. So they
- 18:15just enable and open up this really kinda
- 18:19interesting
- 18:20world. So you've got a little bit of
- 18:21a graphic up on the screen now that
- 18:23kinda talks about and speaks to clients and
- 18:26servers
- 18:27and how they come together.
- 18:29So with MCP,
- 18:31obviously, you need an endpoint or a thing
- 18:33that you can talk to, that you can
- 18:34have the LLM and the ID or the
- 18:37client that's able to recognize MCP servers that
- 18:39it, like, knows how to talk to it
- 18:41and get out there. So a server
- 18:43and the implementation of a server can be
- 18:46local or it can be remote. So for
- 18:49example,
- 18:50there is
- 18:51an Azure,
- 18:53there there's a Microsoft Learn MCP server. And
- 18:57what the folks at Microsoft Learn did was
- 18:59they actually stood up an MCP endpoint, and
- 19:02it's an endpoint that's compliant with the MCP
- 19:04protocol.
- 19:06And you can install you can install an
- 19:09MCP server for Microsoft Learn into
- 19:12your local client. But when you're installing that
- 19:14server, all it's really doing is talking to
- 19:16a remote HTTP endpoint. It's doing, like, SSE
- 19:19and stream mobile HTTP and things like that.
- 19:21The alternative way you could do it is
- 19:23and how some of these have manifested is
- 19:25they've manifested as local server implementations.
- 19:28So literally you go in and you say,
- 19:30okay. Here's the configuration for an MCP server.
- 19:33And quite often these things will be spun
- 19:34up inside of, like, containers or just little
- 19:37m MPM servers that are running locally, where
- 19:40it's running a web endpoint
- 19:41locally on the same client where you're executing
- 19:45the client with the prompt and things like
- 19:47that, and you've got a little bit of
- 19:49a different flow there. But you can always
- 19:51mix and match all these things. Right? Like,
- 19:53this is all just, like, how many of
- 19:54these things do I chain together? What what
- 19:56what are they, and how do they all
- 19:59manifest? So the biggest thing you gotta figure
- 20:01out is, like, hey. Like, is there,
- 20:04is is there a use case for me
- 20:05with MCP, you know, given your persona,
- 20:08you you know, your work identity,
- 20:10whatever it is you happen to do and
- 20:11and how you wanna use it? And if
- 20:13yes, then what's the client I wanna use
- 20:15to get there? Because there are a there's
- 20:18a pretty broad set of clients that can
- 20:20leverage MCP integrations.
- 20:22But that said, you gotta kinda, like, pick
- 20:23the right tool for the right job, and
- 20:24and that's a little confusing right now because,
- 20:26like I said, things like Copilot desktop, Microsoft
- 20:30three sixty five Copilot, they don't have the
- 20:31ability to integrate with an MCP server, nor
- 20:34does the, I believe, the OpenAI desktop client.
- 20:37But that said, the Claw desktop client does,
- 20:39the perplexity client does, Versus Code does, Cursor
- 20:42does, WinSurf does, like, like, all these different
- 20:44things that are out there. And then and
- 20:46then, like, I'm I'm missing a whole list.
- 20:47Like, if you go on, like, the anthropic
- 20:48documentation, like, they've got, like, gobs and gobs
- 20:51and gobs of clients and which ones work
- 20:54with resources, which ones work with prompts, which
- 20:56ones work with tools, all those kinds of
- 20:58things along the way. I would imagine for
- 21:01most of the folks that probably listen to
- 21:03our podcast who are living in the
- 21:05Microsoft ecosystem,
- 21:07you probably work with an MCP
- 21:10capable client already. I know, like, I live
- 21:13in Versus Code, like, 90% of my day.
- 21:16That's all just there and ready to go
- 21:18and
- 21:19available to me. This stuff is, like it
- 21:22looks a little geeky, like, when you start
- 21:23to get into it, but, you know, you
- 21:25need to know, like, hey. How do I
- 21:26define an MCP server that already exists out
- 21:29there? And some of the clients, you just
- 21:30have to pull like, just post, like, and
- 21:33copy and paste, like, raw JSON into these
- 21:35things to get them going for the configs.
- 21:37So that can be, like, a little intimidating.
- 21:39But once you've done all that, like, really,
- 21:40it just becomes chatting with an LLM, and
- 21:42now you've extended the you've extended the capabilities
- 21:46of that LLM. And by extending the capabilities
- 21:48of that LLM, I think really what we're
- 21:51saying is we're extending the capabilities of ourselves,
- 21:54right, because we were using those LLMs to
- 21:56enable our jobs and enable our workflows and
- 21:59and move those things forward. So that that's
- 22:00kind of an exciting thing. For sure. And
- 22:02like you said, there's a whole list of
- 22:04clients.
- 22:05I think OpenAI is getting closer to having
- 22:08more. Like, with some of the connectors they've
- 22:11put in there, their desktop client is getting
- 22:14close, and I've started playing with a few
- 22:16of these clients.
- 22:17Because of that, with Claude, with OpenAI, with
- 22:20Versus Code,
- 22:21all of those. But one thing that's interesting,
- 22:23and this is something again, we're talking about
- 22:26using these clients, hooking them up to other
- 22:28data sources, pulling them into an LLM, kinda
- 22:31like this architecture diagram
- 22:33you were talking about where you have your
- 22:34MCP host or your AI application, OpenAI Cloud.
- 22:37You're using the clients, and you're going out
- 22:39and connecting it to all this data. Tying
- 22:40that into another concept you mentioned where
- 22:44Anthropic described this as the like a USB
- 22:47c port where you can have unlimited USB
- 22:49c ports plugged it in. What is the
- 22:51first thing you you're taught with data security,
- 22:53Scott, with USB ports? You plug in any
- 22:55random stuff you can. Plug them in. Right.
- 22:57When you see a USB port, just plug
- 22:59it in. I think that's one interesting aspect
- 23:01too when you start thinking about these MCPs
- 23:03is to that extent, what are you plugging
- 23:05your data into? If I go in and
- 23:07you mentioned
- 23:08you give it access to your calendar, you
- 23:10give it access to your Notion database, maybe
- 23:13you're giving it access to a SQL server.
- 23:14You're giving it access to your file server.
- 23:16It's actually now pulling data from all these
- 23:19different data sources
- 23:20into an LLM. I know SharePoint is one
- 23:23that has started coming up more and more
- 23:25as well is
- 23:27from that security side, like Microsoft three sixty
- 23:29five, you can do sensitivity labels, you can
- 23:31do security.
- 23:32They've built a bunch of different controls into
- 23:35SharePoint
- 23:36so that you can control how
- 23:39Copilot and
- 23:41the LLMs behind
- 23:43Microsoft three sixty five Copilot
- 23:45interact with your data. What data they're allowed
- 23:47to interact with? How does that handle
- 23:49sensitive information? How does it handle Social Security
- 23:52numbers that it may come across? How can
- 23:55you define different controls? But now what happens
- 23:57if you have
- 23:59a connector, an MCP, that maybe goes and
- 24:02looks at your Microsoft three sixty five data
- 24:06and starts pulling all these files and documents
- 24:09in because there's an API for them. You
- 24:11can get them. But that MCP or that
- 24:13connector
- 24:14may not have all those same controls that
- 24:18Copilot has in place for, especially, I think,
- 24:20of things like sensitivity labels and how you
- 24:22can filter things out there. Permissions, the ACLs
- 24:25that are maybe on some of these files
- 24:26are a little bit easier. But I think
- 24:28you really also need to start thinking about
- 24:30now from a security perspective,
- 24:33whether it's at a corporate data level or
- 24:35even your personal data, what's happening to your
- 24:38data
- 24:39as
- 24:39you're bringing it into these different LLMs? For
- 24:42sure. And I'll I'll I'll throw you one
- 24:43more on top of there. So,
- 24:45if you think about a
- 24:48local MCP server, so like I said, like,
- 24:50often you can spin these things up Yep.
- 24:51Just like a simple m MPM Docker on
- 24:54your machine. Yep. Do you do it in
- 24:57Docker where it's isolated and you have maybe
- 24:59some more of those operational controls, but then
- 25:01you might have to contend with things like
- 25:03container networking
- 25:05and and routing and other constructs?
- 25:07Do you end up in a place where
- 25:09you have just a bunch of random web
- 25:11servers running, like, locally on on your on
- 25:13your machine? Right? Like, how many tie how
- 25:15many MPM servers do you wanna spin up
- 25:17in the background
- 25:18for API endpoints for these specific pieces of
- 25:21functionality?
- 25:22Because really, like, in some cases, what we're
- 25:24talking about sometimes is even, like, little helpers
- 25:28to to do things. Right? Because you could
- 25:29have, like, you could have an MCP interaction
- 25:32where the MCP is interacting with, like, a
- 25:34local file on your desktop,
- 25:36and then it's taking information from that file,
- 25:38pushing it to a cloudy service, a sassy
- 25:40service, whatever that is, vice versa. So I
- 25:43I there there are a whole bunch of
- 25:44considerations there. You know, these things are not
- 25:48super mature in the sense that, you know,
- 25:50they've been out for years and years and
- 25:51years, and we understand all the edge cases
- 25:53and the flows and, you know, do they
- 25:55have the right operational controls, things like that.
- 25:57So it's like, yeah, I think you need
- 25:59to be a little bit, like, careful with
- 26:01them, but I don't know that, like, where
- 26:04things are today
- 26:06that it's as big of a concern as
- 26:08it might be. Like, I see a lot
- 26:09of these things as
- 26:11accelerators for developer,
- 26:13workflows, accelerators
- 26:15for, like, no code, low code kind of
- 26:18workflows, those kinds of things.
- 26:20And, you know, if you're scared of the
- 26:22destructive nature of something, like, just don't use
- 26:24it. Keep doing it the way you've you've
- 26:26been doing it. Right? Like, nobody's saying, like,
- 26:28hey. You have to do these things. But
- 26:30certainly, yeah, like, treat them safely, right, if
- 26:32you can. Run-in a sandbox the first time.
- 26:34Figure out, like, hey, is is is this
- 26:35the right thing for me, my workflow, and
- 26:38what I'm trying to
- 26:40accomplish right now? Yeah. And I think there
- 26:43are some scenarios I think about when we
- 26:45get into some of the examples and start
- 26:47talking about maybe
- 26:48how we've used them that I think about
- 26:50a little bit. There's an aspect to some
- 26:52of these too, depending on the MCP you're
- 26:54using. Like, I would love to see these,
- 26:56and I'm hoping they will get there eventually
- 26:59come to
- 27:00something like Copilot so I can tie some
- 27:02of these MCPs that I wanna use into
- 27:06my Microsoft three sixty five environment so that
- 27:08I can start maybe supplementing Microsoft three sixty
- 27:11five Copilot with some of the data that
- 27:13would come from an MCP. Right now, I
- 27:15feel like I'm in the opposite boat where
- 27:17things like Cloud and OpenAI are building all
- 27:20these connectors and allowing me to pull data
- 27:22in from all of those, including from SharePoint,
- 27:25where I'm starting to sometimes even find myself,
- 27:28it's like, well, it's almost easier to use
- 27:30OpenAI
- 27:31or
- 27:32Claude
- 27:33or some of these other LLMs
- 27:35just because
- 27:37the openness is there to get
- 27:39some of this other data in from other
- 27:41tools that I use
- 27:43where I mean, personally, I would love to
- 27:45just have it in Copilot. Some of it,
- 27:47again, due to the nature of the data,
- 27:48some of it is, like, a lot of
- 27:50these MCPs yet, they're still paywalled. I'm paying
- 27:53for Claude right now. I'm playing for OpenAI
- 27:55right now, and I'm paying for Copilot right
- 27:57now. Just so I can test all this
- 27:58out where if there started to be some
- 28:00of that feature parity, I could not pay
- 28:03for all three of them.
- 28:05Yeah.
- 28:06I think it is tough to get there,
- 28:08right? Like, there's certainly, like, an ecosystem
- 28:10kind of thing. I I will say, like,
- 28:12you can get by a lot with a
- 28:13lot of the stuff just, like, on the
- 28:14on the on the free side.
- 28:16So, you know, like, if you're using, like,
- 28:17Claude, Claude desktop, Perplexity, things like that, you
- 28:21don't need to be on, like, the paid
- 28:22plan to integrate an MCP server with with
- 28:25Claude. I think most of these even like,
- 28:27all all the ones that I can think
- 28:28of, like, they do have some free version.
- 28:31Even, like, GitHub, like, hey. I'm gonna turn
- 28:33on GitHub Copilot because I need GitHub Copilot
- 28:35to be able to do the chat with
- 28:36the LLMs,
- 28:37and then I'm, like, adding these agentic flows
- 28:39to it. That stuff's available and ready to
- 28:41go, and and you can do that pretty
- 28:44turnkey. So you might be limited in some
- 28:46functionality and some other things that are out
- 28:47there. I don't think that stops folks from
- 28:50trying to get, you know, hands on with
- 28:52it and and see where some of that
- 28:54value is. And as we get into
- 28:56our later conversation about, like, maybe some specific
- 28:59MCP servers, at least everything I plan to
- 29:01talk about, like, it's all free or it's
- 29:04already integrated into the ecosystems that you're probably
- 29:06in anyway. You're paying for, like, the underlying
- 29:08thing. Right? Like, if you use, like, the
- 29:09Azure MCP server, then you're probably already an
- 29:12Azure customer. Yeah. It'll be interesting to see
- 29:14where this goes, see where the architecture goes,
- 29:16see how these different clients start bringing them
- 29:19in, incorporating all of it together. Yeah. So
- 29:22why don't we,
- 29:23wrap this one, and then we'll come back
- 29:26for a part two, and and we'll talk
- 29:28about some
- 29:29specific
- 29:30MCP servers and kinda like how we're using
- 29:33them in in our day to day and
- 29:34and kinda how our journey has been here
- 29:36a little bit, and maybe that informs some
- 29:38others. You know, if anybody's out there and
- 29:40you're using an MCP server, like, we we'd
- 29:43love to hear from you, like, hey, What
- 29:45are you using it for? What's going on?
- 29:47Like, what's your specific
- 29:49use case use cases?
- 29:51You know, I I think Ben and I
- 29:52are, like, pretty big on just, experimentation. Ben
- 29:55always talks about his list that is ever
- 29:57growing. I've got a similar one. Like, we'd
- 29:59love to hear how others are enabling, like,
- 30:01their workflows. For sure. So reach out. Let
- 30:04us know.
- 30:05Again, LinkedIn is probably the best. I am
- 30:07finding that as becoming my social network
- 30:10of
- 30:11choice lately Mhmm. Or contact page on the
- 30:14website, mscloudi2pro.com.
- 30:16We'd love to hear from you. Alright. Sounds
- 30:18good. Hopefully, this was enthralling and folks come
- 30:21back for part two.
- 30:22Absolutely. Alright. Thanks, Ben. Yep. Thanks, Scott.
- 30:27If you enjoyed the podcast, go leave us
- 30:29a five star rating in iTunes. It helps
- 30:31to get the word out so more IT
- 30:33pros can learn about Office three sixty five
- 30:35and Azure.
- 30:36If you have any questions you want us
- 30:38to address on the show, or feedback about
- 30:40the show, feel free to reach out via
- 30:43our website, Twitter, or Facebook.
- 30:45Thanks again for listening, and have a great
- 30:47day.