← Model Context Protocol (MCP) for Beginners

🚀 FINALLY! AI That Actually WORKS With Your Apps | MCP Explained in 20 Minutes

Name: 🚀 FINALLY! AI That Actually WORKS With Your Apps | MCP Explained in 20 Minutes
Uploaded: 2025-06-15T09:15:50Z
Duration: 25 min 20 s
Description: 🔥 Stop fighting with AI that can't access your data!

Video 2 of 2 · 25:20

Transcript

Auto generated by YouTube. Click any timestamp to jump to that moment.

Show

0:02topic in the tech community
0:04now. If you see people will be
0:06about this left, right and
0:07This effectively what is model
0:10protocol right? So if you think
0:12the evolution of chat GPT claude
0:15all of these LLMs they are trained
0:17very relevant data over a over a
0:20of time and there's lot of human
0:23thanks to people like us from
0:25who uploaded millions of our
0:26into chat GBT due to the Gibly
0:29these models have become more
0:31um over a period of time but
0:33what is the fundamental problem
0:35is the value of these models is
0:37going just going to be a fancy tool
0:39will sit out of a browser
0:41you start interacting asking with it
0:43be happy about very intelligent
0:45it gives or is there a scope for
0:48to actually do some actions in the
0:50world and that's where anthropic
0:52is the parent company of claude um
0:55some very interesting thought
0:56they wanted these models to have
0:58create some impact in the real world
1:00they wanted these models to take
1:02they wanted these models to
1:05your productivity um to do tasks
1:08your behalf or under your guidance.
1:11was the thought process? They said,
1:12now how can I enable my models to
1:16taking actions? How does it look
1:18How do I create that entire
1:19And they came up with
1:21called model context protocol.
1:23now at a high level, what is model
1:25protocol? It's a way for these
1:28to access um data and tools.
1:31When I say data, let's say your
1:33is sitting out of different
1:35Um like Naven was displaying
1:38dashboard uh just now your data
1:40be in some data warehouses,
1:42etc. Right now the thought
1:46is can all these databases be
1:49in a safe and secure way
1:51LLMs can easily access. Obviously
1:54need to pass some authentication
1:56to tell that this is Rajes
1:58for my own data but then
2:01say the database talks back to you
2:03gives you all the data in real time
2:05care of all your security
2:06a very secure pipeline so that
2:09can use the data to perform perform
2:11of other activities then my data
2:14I I work with lot of tools
2:16I work with as a marketing manager
2:18work with a CRM system like Salesforce
2:21HubSpot where lot of leads lot of
2:24are there left right and center
2:26I have my team sitting out of uh
2:28then I have google sheets excel
2:31sitting out of let's say
2:33sheets and my there are multiple
2:37where I have to make
2:38very simp for very very simple
2:40and those are called tools those
2:42basically applications that you plug
2:44so effectively model context protocol
2:47this superpower to LLMs like
2:50openai manners or deepseek um
2:54a very open protocol so it's not
2:56to anthropic so any LLM can now
2:59talking to these databases talking
3:01these applications and make actions
3:04so that um they they come out of the
3:07web interface that you see and they
3:10perform actions on your behalf right
3:12a high level that is model cont model
3:15protocol right a very very fancy
3:17for um integrations in a very
3:20way So uh I want I don't want
3:23scare you guys but again I what I
3:26to give you is a flavor of the
3:28terms that anyone discusses when
3:30talk about model context protocol.
3:32want you guys to be aware of those so
3:34you know you can you can learn
3:36today. Right. I'll touch base
3:39you can dig into 10th level of
3:41I'll probably stay at level one
3:43level two so that I keep it relevant
3:45a lot of lot of the folks out here.
3:48while I I try to explain the core
3:50of what model context protocol
3:52and then I'll go on to show you a
3:54of demos so that you can
3:56to what I'm talking about
3:57is a theory unless and until you
3:59it in practice right so we'll cover
4:01aspects now you'll hear this term
4:04a lot of times right whenever
4:05talks about model context
4:07they'll talk about n into n
4:11m is getting converted to n plus m
4:13sounds like very crazy
4:15terms some very beautiful
4:18running in the back end. But
4:20does it really mean? Effectively,
4:21they're saying is think about your
4:23organization, right? There there are
4:25departments. Again, the use
4:27is spread across multiple
4:29all across the world. Just
4:31keep the topic relevant and so that
4:34can connect to it. Let's just think
4:36your own company. Let's say you
4:38these five departments, right?
4:40sales, support, analytics and
4:43And all these departments let's say
4:45in silos which they typically do
4:48a marketing um department wants to
4:51with couple of your CRM systems
4:53I need to get the customer
4:56what is the status I need to
4:58to social media um reports or
5:01because I need I have to
5:03what's trending what's not
5:05which campaign is doing good
5:07campaign is doing ugly. I
5:08have to connect to a couple
5:10email services so that I know how my
5:13campaigns are performing or if I
5:15to trigger something right
5:16if you look at um let's say
5:19I need to connect to CRM I have to
5:21to a couple of internal
5:22um and maybe I have to connect
5:25social media etc right so there's lot
5:27redundancy there where every
5:29is trying to connect to
5:32repeated databases several
5:34applications which I need to
5:37with Right? And in a worst
5:39scenario, maybe each of these
5:41want to talk to all these
5:44different databases or applications
5:46which case the network graph or your
5:49data calls will look something like
5:51right? Each of them build some
5:53elements in silos. They want
5:55talk with same stuff uh same database
5:58you know they are building something
5:59silo to talk to that particular
6:01That's what really means when
6:03say uh five n into m in this case n
6:06five 5 into 6 which is 30 complex
6:09This is very similar to uh
6:12say 10 years back where you had
6:14devices in multiple different
6:17right? You had USB type A, USB
6:20B, USB type-C ports and now you you
6:23even know which direction to put
6:25one because one end is different
6:27the other end is different port.
6:29lot of these complexities. What that
6:31is if you have a camera, if you
6:33a mobile, if you have a um let's
6:36iPad and couple of other devices,
6:38need to carry these five to six
6:41plugs along with you just to
6:43each device gets connected to the
6:44socket, right? All these operate
6:48different USB types. So, it's very
6:50Now, what is MCP? Right? This
6:52where N2M consolidates to N plusM.
6:56is a protocol. Effectively it's a
6:59between these two elements
7:01on one side databases and
7:04on the other side and it's also
7:07of condenses all this into a common
7:09right and in our USB example
7:13is what anthropic co keeps on
7:15right MCP is that USB type-C
7:18um in the AI world right now what
7:21it mean so if you look at all your
7:23right now um most of the devices
7:25USB type-C protocol be Beat your
7:28beat your um even iPhones were
7:30to thanks to European laws. Um
7:33iPhone, your Android devices,
7:35has a common port now which is
7:36type-C. So you no longer need to
7:39those three, four different power
7:40with five, six different types
7:42cables. Figuring out which side to
7:44where, plug in, all that headache
7:47gone, right? It's all uniform. It's
7:49single. You buy just a single
7:51single wire, single cable. You
7:54use it front to back, back to front.
7:56that will work with all your devices.
7:59that's exactly what they have done.
8:01they have built a middleware USB
8:04moment uh which basically cuts
8:07all your connections building
8:09things across multiple
8:11uh to a simple seamless
8:14You may ask me hey like this
8:16good theoretically like what does
8:17really mean in terms of
8:19right what this means is
8:22MCPS I think people used to
8:25with something called as uh
8:27right APIs are a way in which
8:30talk to one another in this
8:31case let's say Instagram has
8:33API for instance right um let's say
8:36has an API now what that means
8:38if I have to talk to Instagram based
8:41Instagram's API I will create some um
8:44functions and I'll start
8:46calling Instagram for either
8:48the data from Instagram or
8:50the data or publishing my social
8:52content onto Instagram. Now think
8:54there are there is company 1,
8:562, company 3, company four,
8:58five. All these different
9:00have to create their own
9:01limits or functions in order
9:03call Instagram right effectively to
9:06the same functions which is
9:07the content or pulling
9:10from Instagram. Now these all
9:12into MCP right you just have to
9:15a MCP Instagram server obviously
9:20with some authentication so that
9:22knows who is calling which handle it
9:24want to manage and things like those
9:27your job is done you no longer need
9:29dig into huge API documents and do
9:32this jugard right also what does it
9:35is for developers but what does it
9:37for general audience right who who
9:40even want to go to the API What
9:42means is your claude and GPTs
9:45whatever you interact on the
9:47it only gives answers over there
9:49now you can actually connect it to
9:52something uh to to this Instagram MCP
9:55let's say um and connection
9:58is very easy to do as well and
10:01can manage all your Instagram
10:02out of your cloud desktop right
10:05need not log to multiple different
10:07you can just talk to your cloud
10:09and say hey push
10:11this content to Instagram or
10:13this blog post or create a new
10:16post and it should be able to
10:18it right there from your cloud
10:19desktop on your on your MacBook
10:22that's the beauty of the entire MCP
10:24Any questions on this? Okay,
10:26good. Again, so I'm I'm just
10:27to cover all the key concepts
10:29are being covered in um in the
10:32or model context protocol so
10:34you guys are aware of what it is.
10:36keep it at a very as I told one
10:38two level. So don't worry too much
10:40you don't understand some of these
10:42when you go back dig
10:44read the documents we'll share
10:46documents with you read those
10:48or upload those documents to
10:49GPT or claude and start asking
10:51questions uh that's how typically I
10:53learn lot of these technical
10:55but don't be scared on
10:57at these terminologies right so
10:58does this mean what what are all
11:00components right you'll keep
11:02these terminologies you'll see
11:04talk about something called MCP
11:06MCP client and MCP server server.
11:10are the three main actors in in
11:12typical MCP architecture and there
11:15something called MCP protocol. Right?
11:17you at this uh I mean to at a
11:19level enough to understand that
11:22host plus MCP client this entire
11:25is nothing but your cloud
11:26right instead of web terminal
11:30can install cloud desktop on your
11:31and Windows and that is
11:34your cloud desktop right which
11:36of MCP host and client. the way
11:39talk to different applications and
11:41um is nothing but MCP model
11:45protocol right there's a
11:46way in which they have to
11:48interact with each other there is a
11:50that is defined similar to how
11:52a USB type-C is defined think about
11:55like the wire that goes from your
11:57to the socket uh and how the
11:59is transferred right through the
12:01um that's what MCP protocol is or
12:05communication standard is and uh
12:08are the servers, the databases and
12:10we talked about, right? MCP
12:12are nothing but uh the
12:14or the databases. So for
12:17in the example I gave you, if
12:19have to connect to a connect to Slack,
12:21I'll have a Slack MCP server. If
12:24have to connect to Razer Pay because I
12:26to check it check the payment
12:28get some information from it,
12:30called Razer Pay MCP server, right?
12:33I have to connect to um let's say
12:35or a CRM system, it would be
12:38Salesforce MCP server. Right? So
12:40application and every database um
12:43I'll also show you database which is
12:45we connected to superbase. So that
12:48be a superbase server or superbase
12:50server. Right? So again long story
12:52you have the user which is you.
12:55open your cloud desktop. You start
12:57on your um cloud desktop. you
13:01with some of these MCP servers
13:03right so there's a way in
13:05you connect once you connect all
13:07available for your cloud desktop
13:09is your host and again just to
13:12touch what happens right what
13:13at runtime when you type a query
13:15enough to understand there are two
13:17constructs here MCP host and
13:19server and they talk in in some way
13:22with one another there are uh three
13:25topics within MCP right uh there is
13:28called tools there are
13:30And there are prompts. Now
13:32are these tools? Tools are nothing
13:34these applications we talked about.
13:36say in this scenario we talked
13:38a Slack MCP server. A Slack MCP
13:41might have multiple tools uh
13:43it. Um one tool probably is to
13:47all your uh Slack channel names for
13:49right another tool is to write
13:52or write a message to one
13:54channel. Another tool is to
13:56get all your contacts from
13:58So tools are nothing but what are
14:01individual capabilities within
14:04MCP server that particular
14:06can do and which is exposed
14:08your um cloud right let's say you you
14:12hey claude why don't you send my
14:14sales report summarize it top
14:16key points to my slack uh to this
14:19channel let's say my sales
14:21um then claude understands as a
14:24that hey this guy is talking about
14:27Now let me go to Slack MCP server
14:30you already connected. Nothing to
14:32from your side. The intern I'm just
14:34the internal mechanics. The
14:36server will have all these tools
14:38and it realizes okay this guy
14:40me to send a message first. He
14:42to identify the summary. So it'll
14:44try to get the summary of all the
14:47three clients. Then it'll go to
14:49It'll in the Slack MCP server.
14:52figure out okay there is a send u
14:55to a channel tool that is
14:57to me let me use the tool to
15:00this message on this slack right um
15:02that's how it it closes the loop now
15:05is resources right uh in the same
15:07we said it has to get some sales
15:09and then then send to slack which
15:11I need access to this resource
15:13resource can be either a database
15:14can be a document sitting somewhere
15:17and u resources basically talk
15:19how how you um extract that
15:22typically and as a protocol
15:25gives all this control in the
15:26hand which means it asks lot of
15:29permissions before it can
15:30any document that you request it
15:33access right so that way uh it they
15:36kept some constraints to ensure it
15:38go rogue and sort of do all
15:40stuff and um access all your
15:42they have kept controls in
15:45to ensure it behaves within your
15:47or battles
15:50finally what is prompts right
15:52as you are aware um typically
15:55define how the system should react
15:58in this case MCP's context they have
16:01one level deeper and what they said
16:03hey you know what if you let's say if
16:06are repeating some tasks again and
16:08and you want some standardization
16:10the output you can predefine some of
16:12uh prompts so that your output is
16:16standardized every time you run
16:18you need not enter these prompts
16:19and again uh let's say I asks for
16:21summarized report of my latest sales
16:24the last month. Typically, as Naven
16:26shown, you have to figure out what
16:28prompt is. You have to enter that
16:30and then it extracts all the data
16:32displays to you. But let's say there
16:34four people in the organization and
16:36different people will write
16:37prompts and outputs will be
16:39different. Right? In this
16:40what you can do is you can say,
16:42this guy always wants to analyze
16:44sales. So, that prompt is
16:47into the workflow. All you
16:50to do is slash command analyze
16:52sales and it'll actually pick
16:54prompt and give give you the
16:55Right? So it's a way of
16:57workflows so that your out
16:59whatever you expect from the from
17:02the tool is completely standardized.
17:05this is one example of a workflow um
17:08this is typically maybe I'll show
17:11in a better way in the demo but this
17:13what happens right user will request
17:15what are our top customer issues and
17:18should we do about them. So it will
17:20all the support tickets. Let's say
17:22there is Jira or there is some tool
17:24all these tickets are. Jira is MCP
17:27and tool uh tool is get tickets
17:30that MCP server. So it'll go to
17:32MCP server. It'll make a call this
17:34call. It'll get all the results.
17:37it'll analyze the sentiments. It'll
17:40to identify the patterns. It it'll
17:42some action items. Okay. So based
17:44this let's say I figured out uh my
17:47times are always very bad in
17:50and I have to and a lot of
17:52are complaining about read all
17:54customer tickets it'll understand
17:55feedback it'll categorize the issues
17:58multiple different buckets and
18:00it'll create tasks for
18:01teams and notify those teams
18:03all this is possible with a single
18:06once you connect all the relevant
18:07servers right and in terms of
18:09I think this is what it does
18:11there are there were three
18:12issues identified There are
18:14action items created. Engineering,
18:16support teams, all are notified
18:18Slack. And here is overall action
18:20It saves into let's say ASA where
18:23can track what's happening with
18:25to uh this entire activity.
18:27So it's no longer a siloed
18:30you sitting uh in front of
18:33cloud desktop when you connect all
18:35different MCP servers you are
18:38this superpower uh to your LLM to
18:41take actions on your behalf in
18:43very controlled way. All right. So I
18:45these are some additional
18:46right? So your types of
18:48can be different. It can be
18:50live databases. It can be just a
18:52document that you upload. It can
18:55your system monitoring f files or
18:57data sets. It can be
18:59Again, I think different
19:01of resources that you can get in.
19:03I'll take an example of this. So I'll
19:06forward. And typically just
19:08on what really happens um
19:11used to happen before and what can
19:13going forward. This is an example
19:14the prompt where let's say if you try
19:16analyze there are inconsistent
19:18there are you cannot get the
19:20results every time as we discussed
19:21MCP prompt is a good way to get
19:25standardized results out out of the
19:27so what is the business impact
19:28so what's in it for you why should
19:30even care about MCP right as I told
19:33the best standard for AI
19:34giving superpowers to your
19:36improve your productivity by huge
19:39reduce time spent on redundant
19:41because now the LLM has all
19:43right power and ammunition to go
19:45it. Uh there is no lock in
19:47because it's not limited
19:49anthropic or cloud per se. You can
19:51any LLM in the world. It's a open
19:54So you can build applications
19:56top of it. You can move from claw to
19:59GP or a deepse right. All of these
20:01follow the same USBC protocol or
20:04context protocol. So everything
20:06work seamlessly seamlessly
20:08of the LLM that you choose.
20:10I think uh that is overall MCP for
20:14in a nutshell. Any questions guys
20:17I jump on to demos? Okay, so
20:19has a question. Fancy PPD, how is
20:21made? I'm assuming using some AI.
20:23I basically used uh we are working on
20:26coding agent. Uh so this was made
20:28that AI coding agent. Again there
20:30some interesting tools like gamma
20:32which you can make but I wanted to
20:34some visualizations like these how
20:36MCP goes through different tasks and
20:39some visual animations of the nodes
20:41all. So I I used our local AI coding
20:45to do this still in early phases
20:48we will most likely launch it next
20:50Yes, I'll jump into demo mode
20:52There's one more question. Yeah,
20:54in the chat. Is MCP kind of agentic
20:56which can do tasks automatically by
20:59with AI agents? It's asked by
21:00Manoharan.
21:02Wes, you're right. MCP is more of
21:05I mean model context protocol, right?
21:07like it's a protocol that is
21:10At the end of the day, it's AI
21:11right? It's letting your LLM
21:14with lot of different data or
21:16to get your action done, right?
21:18a way to connect the data is how I
21:19put it. What you connect to,
21:22to is up to you. Do you want to
21:23to a coding agent? Do you want
21:26connect to a let's say sales AI agent
21:30a market research AI agent? All those
21:32possible. So it's basically a way to
21:35your LLMs by connecting to
21:37tools and data or AI agents
21:41also fall application layers. I
21:43anthropics which has basically
21:45MCP in the current aar is the
21:49and the foremost and it'll
21:52set the benchmark the baseline
21:55how communication should happen
21:57these analogies Rajes
21:58and I think that'll set the
22:01if you will so um like the USB
22:05that is used very
22:06with MCP just for people
22:08understand uh and Rajes gave enough
22:11on Right? That's the easiest
22:13of thinking about it, right? It's
22:15plumbing. It's that condute that
22:17an LLM with the myriad tools
22:19capabilities on the other end and
22:21able to sort of make it very very
22:24down if you will by still being
22:26to interact once you build the
22:28primitive connections to be able
22:30get the system working in as
22:32a manner as possible. Yeah.
22:34I think guys uh these terms sound
22:36fancy and crazy and I think when
22:39look at LinkedIn, Twitter people
22:41about lot of things um without
22:44deep. So there lot of technical
22:46that is thrown left right and
22:48So don't worry about any of
22:50It's very simple giving
22:52to LLM LLM so it can take
22:54That's MCP for you in in a uh
22:58Uh and I think what you guys
23:00do probably I'll I'll run walk
23:02through demo also real quick but
23:04you should guys should really do is
23:05exploring and playing hands-on
23:07that's when you'll truly
23:09the unlock that this uh
23:11provides and you'll start appre
23:13it lot of big players have
23:16coming in onto the MCP uh very very
23:19I think uh razor pay postman any
23:23tech company that you see everyone
23:24launching their MCP servers so that
23:26they can supercharge their workflows
23:28their particular applications. Uh
23:31are very interesting use cases
23:33out of MCP. Um for instance,
23:35to your app store where you can
23:37all the applications and you can uh
23:39to each of them. I think this
23:41was created almost a decade
23:43But now people are think people are
23:45that this MCP servers um
23:49uh something called MCP server
23:51can itself become like as big as
23:54app store where you become
23:55platform uh for all the
23:57servers also it's slightly at an
24:00stage though evolving very very
24:02where connecting to these MCP
24:05feels slightly tricky today I
24:08say very difficult but slightly
24:10so people are coming with
24:11constructs on how to connect
24:13these MCP servers. You'll keep
24:16terms like okay local server,
24:18server, blah blah blah all those.
24:20don't don't uh really worry about all
24:23of them at this particular stage. Uh
24:25try to connect to some pre-existing
24:28MCP servers. I'll show you a couple
24:30them. Start playing with them as
24:32one. In phase two, you can let's
24:34you have some redundant activity
24:36you perform on a daily basis. You
24:38yourself create an MCP server for
24:40use case. Um that is probably the
24:43step. And the third level is
24:45multiple servers. See how u the
24:47happens like how your LLM
24:50with all of them with single
24:52instructions, right? Without you
24:54into very detailed instructions.
24:55uh take it stepwise. Don't jump into
24:58three without checking out what
25:00need to do on phase one. But again
25:02one step at a time. Don't fall into
25:04trap of technical jarens. Get
25:07Start implementing. Start
25:09the results by yourself. All
25:11I'll switch gears. I'll jump onto
25:13demos now.

⚡ How to Add ANY MCP Server to Claude Desktop | Basic Tutorial

Want the next one in your inbox?

Join 1,000+ Product Managers getting one deep dive every Friday.