I learned MCP for you. In the past 68

months, you've probably heard a lot of

people talk about MCP. MCP is so

amazing. MCP is such a game changer.

Well, I must tell you, they are correct.

I have taken a bunch of courses, build

MCP servers and clients for our team and

a couple B2B clients that we work with

as well. We've also included a section

on MCP in our AI agents boot camp. So,

in this video, I'm going to explain to

you the fundamentals of MCP, show you

how to use MCP servers and how to build

your own MCP servers as well, both with

code and no code. Throughout this video,

there will be little assessments, which

if you can answer these questions, then

congratulations, you'll be educated on

MCP. Now, without further ado, let's go.

A portion of this video is sponsored by

Balt. Okay, a quick overview of the

structure of today's video. First, we're

going to define MCP, talk about why it's

important, and how it is that we use

them. Then we'll be diving under the

hood in which I'll cover the

fundamentals of MCP, the concepts and

frameworks that govern it. Basically,

everything you need to know to be able

to use and build your own servers. But

it is not enough just to know all these

concepts. So I'll actually show you how

to build one as well, both with code and

no code. Let us now define MCP. MCP,

which is the abbreviation of model

context protocol, is defined by

Enthropic, the people who made it, as an

open protocol that standardizes how your

LM applications connect to and work with

your tools and data sources. Okay. What?

So first time I read that definition I

was like sounds important. So let me

actually explain it to you the way that

I understood it using an analogy. MCP is

like this USB plug here. It allows any

device to be able to connect to any

other devices to be able to work

together. Prior to this universal

standard of the USB port, if you wanted

to connect things to each other, like

say your computer to your microphone, to

your webcam, you would be having a very

bad time because all of the ports to

connect with each other would be

different, which means you would have to

get a bunch of different special cords

and different adapters to manage to get

them to work together. And then you

would have to write the software to

cater towards like those specific port

thingies, etc. Very complex, very messy.

But with this USB port, we're able to

now just connect all these devices to

each other and they just work. Amazing.

Wow. Do not be fooled. It sounds like

such a simple little thing, right? Just

like standardizing the port connectors,

but it amounted to massive amounts of

innovation with this ease of use. Now

before MCP existed, if you have your

agent and you wanted to give it access

to different tools and different context

or different types of data, you would

have actually faced a very similar

problem to the pre- USB port days

because each external system and tool

that you're trying to connect to has

their own specific definition of how it

is that you interact with it, which is

their special API. For example, if

you're trying to make a scheduling

agent, you would need to give it access

to like your calendar, your email, a

note-taking app, Zoom, Calendarly, and

you would need to custom write the code

for each specific tool that you want to

give your agent, which is a pain in the

ass. So that's why when Enthropic

proposed the model context protocol, the

MCP, which would standardize the ways

that we interact with all of these

external softwares, everybody was like,

amazing. Wow, so good. Now your agent is

able to have access to all of these

different external softwares and tools

using the same protocol. Within just

months that the number of tools that you

can give your AI agents now, which are

called MCP servers massively exploded to

thousands. Now there's over 20,000

pre-built MCP servers that you can just

grab and give your AI agent. Here is

just a small sample. Now, anybody can

easily write them, publish them, and use

them. Let me actually show you exactly

how easy it is to use an MCP server.

Like say for example, I'm looking at

these MCP servers here. You can pick any

of them, but let's just pick like Alpha

Vantage for example. It is a MCP server

that enables LLMs and aentic workflows

to seamlessly interact with realtime and

historical stock market data through the

model context protocol. Okay, great. So,

you can get access to stock market data.

You copy this stuff which allows your LM

application to have access to the MCP

server. copy it into your LM application

like for example cloud code if you want

to give the tool to cloud code and voila

you'll be able to have access to the MCP

server and all the different tools that

are defined on the server then you can

do something like plot the coffee stock

market prices for the past 10 years then

Claude will say Claude wants to use the

coffee function from Alpha Vantage are

we going to allow this and we'll be like

okay allow once using that tool was able

to get the coffee price data and now

it's going to created interactive

visualization and there you go, coffee

global prices with all of this

information. Oh my god, coffee is so

expensive these days. So yes, this is

pretty much how easy it is to use any

MCP server now. And I can also very

easily switch what is called the host

which is cloud code in this case to

something else like NA10 for example or

my own custom AI app. So let's now

actually dive into the foundations of

MCP so you will exactly understand what

just happened here and how to build your

own MCP servers as well. And by the way,

MCP servers are actually even more

powerful than just giving your agents

tools. There's a lot more that you can

actually do, which is why people are

able to actually build these MCP servers

and then sell them.

The best course that I found which

covers the foundations of MCP is

Anthropic's own course in collaboration

with deep learning.AI called MCP, build

rich context AI apps with anthropic. The

course starts with explaining basically

why is MCP important. So I already

covered most of that. Um but I do like

the diagrams that they gave which shows

that see before MCP there was fragmented

AI development with each AI app you

would need to have a custom

implementation custom prompt logic

custom tool calls and custom data

access. And you got to do that for all

of them versus with MCP which is a

standardized AI development for each MCP

compatible application you're able to

have access to a lot of different types

of MCP servers. For example, you can

have access to a data store MCP server

so that you can have access to data

stores. You can have a CRM MCP server so

that you can have access to CRM systems

and a version control MCP server so you

can have version control software. All

plug-andplay and the other way around as

well. Every time you build an MCP

server, it would be reusable by various

AI applications like a Google Drive MCP

server that has access to Google Drive

functions that can be used by MCP

compatible AI assistant, MCP compatible

AI agent, and an MCP compatible desktop

app or IDE too. All right. Next

important concept to understand is the

client server architecture. Within the

components that make up MCP, there are

three different components. The host,

client, and server, HCS for short. And

the way that they relate to each other

is the host houses the MCP client, which

goes through the MCP protocol in order

to have access to the MCP server. Hosts

are LM applications that want access to

certain data or tools through MCP. As I

showed earlier, the cloud desktop is an

example of a host. NA10 is also an

example of a host. You can also have

IDs, AI agents. Really any large

language model application can be a

host. Now the host wants to use the MCP

servers which are the lightweight

programs that expose different

capabilities like the servers here like

alpha advantage we talked about earlier

that's able to get real-time historical

stock data. Uh there's also time

provides time and time zone conversion

capabilities. Cherry studio that's used

for LM provider support like AI

assistance and conversions, documented

data processing. Postgress SQL to

provide readonly access to Postgress SQL

databases allowing schema inspection and

execution of readonly queries, GitLab,

Reddist, etc., etc. There is a lot. Now,

in order for the host to actually have

access to these MCP servers, they need

to maintain what is called an MCP

client, which lives inside a host and

evokes the MCP protocol to maintain that

one-on-one connection with the server.

Got it? MCP client lives in MCP host to

invoke the MCP protocol in order to have

access to the MCP servers which the

players being HCS host client and

server. All right, time for a little

quiz for the example that I showed

earlier which allowed cloud desktop able

to access historical stock data alpha

vantage and cloud code. Which is the

host and which is the server? Put your

answer into the comments. Okay, let's

now talk about what are the things that

are actually contained within an MCP

server. Well, there are three major

things. These are tools, resources, and

prompt templates. TRP tools, which are

the things that we've kind of seen the

most of are functions and tools that can

be invoked by the client. So, these are

these things over here like time series,

intraday, time series, daily, blah blah

blah, symbol search, global quote. Also,

stuff like sending a message in Gmail,

calculator, retrieval and search,

sending a message, updating database

records. All of these are tools. But

that is actually not all. You can also

have what are called resources which are

readonly data that is exposed by the

server. So this is the data that the

client is able to query but cannot

change. For example, there are certain

files like markdown notes that's

contained within the server that you

might want to read. Maybe it's tracking

logs. Say for example, every time you're

calling a weather app tool and you're

getting that data, you might want to

actually save a record of all the

different weathers that you've

collected. So at some point you might

want to compile that together and make

like a visualization or something,

right? It would not be very efficient if

every single time you had to go collect

a weather data point. You would have to

like use the tool and spam the tool over

and over and over again. It could also

contain just database records of stuff.

Like for example, there's just like

information that you're storing like

contracts, meeting recordings, and

notes. Then there's also prompt

templates. These are structured prompt

blueprints so that you don't have to

force your user to have to come up with

their own prompts. Like for example, if

you are someone who's trying to use an

MCP server that uh is able to summarize

like a certain transcript from a meeting

and then generate a report based upon

it, you could just write something like,

oh, summarize this transcript into key

action items, then generate a report,

right? Like you could write something

like that, but that probably won't yield

the best results. So instead, the people

that actually built the MCP server can

put within their MCP server a prompt

template that is much more flushed out

that's able to actually extract the

insights and generate the report in a

much better fashion so that the user all

they have to do is input like some

specifics into that template. This takes

the burden of the prompt engineering

from the user. As a full example, you

can have like a SQLite MCP server which

gives access to like a database which is

the SQLite database. Don't worry if you

don't know what that is. is basically

just like a database that can store

information. Anyways, so you can have

tools that allow you to read the SQL

database to insert information into the

database to change the information in

the database and to delete information

from the database. In addition to that,

there is a resource that is logging all

of the changes that are happening to the

database so you have a historical track

record of what's happened. And there's

also some prompt templates that

incorporates the best practices to

interact with that database. Make sense?

All right, time for our next little

assessment. Please answer the questions

on screen now and put it into the

comments so you are following along. MCP

is all about giving your agents more

context so that they can do really

powerful things. So when you're ready to

deploy your AI agents with a lot more

functionalities and you want really good

UI, NA10 plus Bolt is a really powerful

combination and you don't need to write

a single line of code or you can just

directly build via Bolt. Bolt can take

you from an idea to a fully functional

product within minutes. Whether that is

software for your team, an internal tool

or a website. It has everything that you

need, whether that's hosting, backend,

database, off, domains, and even SEO.

One of the things I really like about

Bald is that it seamlessly handles the

entire process so that you can focus on

building real things as opposed to

handling infrastructure. You don't need

to handle like setup, configs, war,

external tools. Bolt puts the most

powerful tools used by developers into

your hands with an intuitive chat

interface. It really is as simple as

this. So, here we have the Bolt new

landing page. And we can just type

something like build an app that blurs

out people's faces in pictures and just

click build now and then let it do its

thing. Of course, you can go into a lot

more details and have a lot of like

specific specs within the app as well.

And look at that. Here we have this

application and let's actually try it

out. We will put an image. Here is a

picture of me when I was still a little

Tina. And there you go. We have the

original and then we have the blurred

version. And you can download the

blurred image as well. As well as you

keep building, you can include a lot

more features and tweak things however

it is that you want. Let Bald do the

heavy lifting so you can focus on your

vision instead of fighting bugs. You can

try out Bald V2 for free today at this

link over here, also linked in

description. Now, back to the video.

Okay, so the final concept that I want

to cover for MCP before diving into

actually building them is the

communication life cycle. The way that

the MCP client and the MCP server

communicate with each other. The first

step is the initialization when the MCP

client interacts with the MCP server

being like, hey, I would like things

from you now and they initiate a

connection. Then you have the message

exchange phase when the MCP client makes

a request to the server. for example

like I want to use this tool and the MCP

server goes like okay and gives it back

which is the response. So you do that

exchange the messages and then finally

there's the termination phase when the

client and the server terminate and stop

interacting with each other. Now the way

that the messages are done is called a

transport. The transport handles the

underlying mechanics of how messages are

sent and received between the client and

the server. And there are two different

categories of how transports work. First

one is when the server is running

locally like your MCP server and your

host all of this is within a single

machine. The analogy that I like to give

here is say that your MCP client is like

the diner and the chef is like your MCP

server who is going to be the one

preparing the food for your diner. Now

for the server and the client running

locally, it'll be like just you and your

friend cooking at home with like a

notebook. You as the chef is just like

cooking it and maybe you're just like

writing down what it is that you're

doing and like just passing a note to

your friend the diner for some reason.

So they're like keeping updated about

what's happening while you're cooking.

That is like the standard out a running

log that's just like living locally

between you guys. You don't need to rely

on like a waiter or a cashier or

anything like that. The other category

of transports is for remote servers. So

these are servers that live in other

machines on the cloud somewhere else.

And there are two different approaches

to interact with a remote server. The

first one is called HTTP plus SSE which

is server sent events. Don't worry so

much about the terminologies. It

supports what is called a stateful

connection. Going back to that analogy,

in this case, you would have to like

leave your house and say you're at a

restaurant. You, as a diner, would sit

down at the restaurant. You're the

client, and you would interact with a

waiter, which is a server, and you would

tell it stuff like, "Oh, here's my

table. Here's the appetizers that I want

to order, blah, blah, blah." And then

your server would like bring you those

things, and say if you want to order

like the same drink again, you could

just say something like, "I'll like to

have the same drink again." And you

don't have to like exactly specify

whatever it is that you ordered the

first time. So this is called a stateful

connection because the server maintains

the state of it like remembering the

knowledge the memory of what you ordered

all this information throughout the

entire interaction. Now there's also

what is called stateless stateful

stateless. This kind of transport

supported by a streamable HTTP. To be

stateless is sort of like if you go to a

fast food restaurant as opposed to like

a sitdown restaurant. Like if you're at

McDonald's for example, you would order

like a Big Mac and I don't know like

fries. But then like if you want to get

something else, you want to also get

like chicken nuggets, you would actually

have to go and like input your order all

over again and get like a new ticket and

you have to wait again. The cashier

that's clashing you out, the server

doesn't actually remember everything

that you ordered previously. So that's

why it's called stateless because it

doesn't remember all of the knowledge

that came with your prior interactions.

Each request is independent of each

other. So without going into way too

much detail about this like the

streamable HTTP is the preferred

transport method because it's able to

support both stateful and stateless and

there are instances in which you want to

have like stateful interactions and

instances in which you want to have

stateless interactions as well. So this

section might sound a little bit like

detailed and theoretical but it does

become very important when we're

actually going to be using our MCP

servers and building our MCP servers

which is right after I give you this

final little assessment.

Okay, I'm going to show you the no code

implementation of building your very own

MCP server now. And we're going to be

doing this by using n10. So you can do

stuff like send an email to Tina with

the address tina@ lonelyotopus.com

asking her if she wants to grab a coffee

sometime.

Execute workflow. Press enter. goes

through this. We can see that it's

accessing the AI agent is accessing the

MCP client which is here. It contains

the ability to have a calculator and

also the tool that allows to send a

message to Gmail. So, it's able to

access the tool to send a message to

Gmail so that you will be able to get

this email. Hi Tina, I hope this message

finds you well. I was wondering if you

like to grab a coffee sometime. Let me

know what works for you. Best regards.

Email was automatically sent with NATO.

Okay. So, how to build this from

scratch? start off with a fresh new

workflow using NA10. If you have no idea

what NA10 is and you're not familiar

with it at all, check out this video

over here. I go into details about how

to build AI agents using NA10. I'm not

going to go into this here, so the video

is not super long, but basically it's a

no code low code tool that allows you to

build workflows. Also allows you to

build MCP servers. So you have your

server trigger over here and you want to

add some tools. Like for example, we can

add the calculator tool, which is just

like the ability to calculate things and

do math. Not super innovative but yes.

And then we can also add the email tool.

So maybe like a Gmail tool. And in here

we need to set up the credentials like

create new credential. You can kind of

go through the process. So all you have

to do is like sign into your Gmail um

which I will do here. And then you'll

have the account that is connected. Yay.

Then um because we just want to send

emails here. So the operation here is

going to be send. So for two we're going

to let the model decide these. So, let

the model decide the subject. Let's al

also let the model decide uh we'll have

the email type HTML and the message.

Let's also let the model decide that as

well. Wonderful. Amazing. So simple.

Right now, let's save this. Now, in

order to actually use our server, we

want to come here, make sure that it's

active. So, the server itself is going

to be active. And we want to take this

production URL. So, this allows access

to the server. Now I have another

workflow over here which is an AI agent

in which you can prompt it directly and

respond also attach the open AI chat

model here. So in order to add the MCP I

can click the MCP client tool the

endpoint we're going to copy paste what

we got from the MCP server and here in

the server transport see we have the

server send events which is deprecated

and the HTTP streamable the two

different transport right so we're going

to click HTTP streamable transport and

we don't need to change anything else

over here. Wow. Now we have the MCP

client. So let's actually test it out.

Let's actually do math this time.

Calculate 10

11 * 99.

We can see here that the AI agent goes

over here. Human says calculate 11 * 99.

It goes to the MCP client and it uses

the tool cal the calculator tool which

we have. And it says useful for getting

the results of a math expression. And

then the open AI chat model was able to

get the results by using the MCP client.

And then we were able to get the final

result here with the output being the

result of 11 multipi by 99 is 1,089.

Yay. See how easy it is. There's a lot

that you can do with this. Of course,

you can add like so many more tools into

your MCP server. And you also don't need

to have NAN as your host. We can easily

switch out the host for something else.

Like for example, maybe you want to

change out the host for cloud desktop.

So what we need to do is go on cloud

desktop. So go on help, type in

developer. So switch to developer mode

and then under developer open the claw

desktop config file. So this should be

empty. You can open it with whatever

text editor. So I'm here I'm just using

VS code for it. And you want to copy

paste this. So this is defined MCP

server and we're using NAN because

that's where the server is and we are

using SSE over here. And over here just

paste the endpoint here. So command S or

control S. Save it. Restart Claude and

wow amazing. We have NA10 over here

under the tools and includes send a

message in Gmail and the calculator. And

then you can do the same thing here.

Send an email to Tina@ lonely

octopus.com

asking her to host an octopus party.

Click enter.

See it'll say over here Claude wants to

use send a message in Gmail from NA10.

And you can click allow once to allow it

to use the tool from the MCP server. And

moment of truth, we have over here, "Hi

Tina, I hope this message finds you

well. I want to reach out to see if

you'd be interested in hosting an

octopus party. I think it'll be a

wonderful event and you'll be the

perfect host for it. Let me know if

you're available and interested. Looking

forward to hearing from you. Best

regards. Yes, automatically sent with

NAT." Voila. See, super easy, right?

Okay. So, now I want to show you the

code implementation of doing this. Now

with code you are able to do a lot more

than this no code approach. I think the

no code approach is amazing. It's great,

right? It's really accessible and easy

to use. Remember we talked about earlier

about having like additional resources

and also prompt templates. You as of

right now cannot do that using NAT. You

can't actually put in the prompt

templates and specify those and you also

can't put in additional resources as

well. And in the code implementation you

can. So I'm not going to go and show you

every single line of code right now

because I don't want to make this video

like super duper long. But I want to

show you like the snippet of these are

the tools that we're going to be using

and this is how you can actually get it

to run. Okay. So for this code demo,

we're going to be using cloud desktop as

the host and I'm going to showcase a

simple MCP server that's able to provide

tools, resources, and prompts for

working with Google Sheets and Google

forms. So we can tell that the MC3

server is connected because if you go to

G Sheets over here, which is the MCP

server name, you can see all the tools

that are contained over here. So I have

a lot of different spreadsheets that are

available. So let's actually ask list

out the spreadsheets that are available.

It will want to use the list

spreadsheets. So this is the tool and it

tells us that we have 20 spreadsheets

that are available in the Google account

like NA10 form, anonymous feedback, the

only octis alumni blah blah blah etc.

All of these different ones. Would you

like me to read the data from any of

these spreadsheets here? I can say use

read sheet to get the columns

of the anonymous

feedback spreadsheet.

So it wants to use read sheet and this

is actually a resource. So it allows you

it's a readonly file where you can

actually read what the column names are.

You can see the spreadsheet ID and the

information that's contained over there.

And finally, I want to show you what it

looks like to use some of the prompt

templates. So, if we click over here,

add from sheets, here are some of the

prompt templates. So, let's click

analyze sheet data. And we can say use

this to analyze the anonymous

pretty spell anonymous

feedback

spreadsheet

and create a dashboard just to make it

look pretty.

And we can see here is what the prompt

template actually looks like. And it

shall do this. Let it do its thing. So

while it's doing that, let me actually

show you what the code looks like and

how to run this MCP server. Here's the

readme. If you follow these

instructions, you'll be able to run this

MCP server yourself. You can see that

the tools that it contains include

things like list spreadsheet, read the

sheet, write the sheet, append the

sheet, etc. It also has the resource

where you can use the sheet in order to

access sheet data as the resource. And

it has prompt templates including

analyze sheet data which is template for

analyzing sheet data. Creating report

template. So a template for creating

report spreadsheets and form to sheet

which is a template for form to sheet

workflow in the code itself. Not going

into too much detail over here. You can

see that after we define the credentials

the way that you can define these tools

that you can decorate the function. So

you know that's MCP tool. So this one is

the function for listing the

spreadsheet. This one is for reading the

sheet etc. And then the way that you can

show that it's a resource is that you

just decorate it so that you know that

it's a resource. So get sheet resource.

And here are the prompt templates

themselves. So they're actually

hardcoded directly in these different

prompt templates. So let's go back to

claude here and see what is happening

there. Great. We see that it analyzed

the anonymous feedback spreadsheet. So I

read that information, analyzed the data

using the template that we provided and

it actually created a dashboard as well.

So here is the dashboard. So total

response is 43. The average MPS score

which is the net promoter score is 4.9

5.98 blah blah blah all these things. So

it tells you information about this

different categories uh how the study

plan progress tracker is qualitative

information about what attracted

students to our boot camp and insights

as well. So different types of

information including what works well

and what doesn't work well etc etc. So

yeah there you go. This is the code demo

of an MCP server. I hope you can see

that there's so much potential for so

many different types of things that you

can build now that you have the ability

to create your own MCP servers. You can

of course also change your host. You use

this MCP server with NA10 uh with your

own agent with your any large language

model application that you build as

well. The possibilities are endless. So

if you do want to dig deeper into this,

I highly recommend that you actually

check out the anthropic deep learning AI

course that goes into the exact

implementation step by step how to build

this entire MCP server. It is really,

really, really powerful. There are of

course a lot of other ways that you can

implement MCP servers as well, both code

and no code approaches. I also really

like this one by Network Chuck where he

goes into a lot of detail about how to

use Docker in order to do this. So I'm

going to link his video below as well. I

really recommend that you check it out

if you're interested in implementing it

using Docker. But for this video, I'm

going to end it here. I hope that you

now have an excellent fundamental

understanding of MCP and how to actually

build them yourself, too. Here is the

final little assessment. Please answer

these questions and put them into the

comments.

Hope you enjoyed this video and thank

you so much for staying until the end.

I'll see you guys in the next video for

live stream.