Model Context Protocol (MCP), clearly explained (why it matters) - YouTube

https://www.youtube.com/watch?v=7j_NE6Pjv-E

Transcript:
Intro
0:00
everyone is talking about mcps it's gone
0:04
completely viral but the reality is most
0:08
people have no idea what mcps are and
0:11
what they mean and what are the startup
0:13
opportunities associated with it so in
0:16
this episode I brought Professor Ross
0:19
Mike who is probably the the best
0:22
explainer of technical Concepts in a
0:25
really easy way that someone who's
0:27
non-technical can really understand I
0:29
brought him on he explains it
0:31
beautifully in such a short amount of
0:33
time and if you stick to the end you'll
0:35
hear a couple of his startup ideas that
0:37
incorporate mcps so um enjoy the episode
0:41
and see you
0:43
[Music]
0:50
soon all right well we got Professor
0:54
Ross Mike on the Pod um and the reason
0:59
why we have him is because I don't know
1:02
what the hell mcps are and I've been
1:04
seeing it on X and I need a succinct
1:09
clear Professor Ross Mike
1:12
explanation um yes I've read a bunch of
1:14
threads on it and I've seen a couple
1:16
videos on it but there's nothing like a
1:18
Ross mic explanation so I'm here for the
1:22
what do I need to know about mcps and
1:26
that's that's what that's why you're
1:27
here thank you for coming on I I I
1:29
appreciate that thank you very much yeah
1:31
class is definitely in session I'll just
1:32
start um sharing my
1:35
screen okay so understanding
1:38
mCP um is
1:41
really um important uh but you'll also
1:45
realize the benefits and why it's sort
1:47
of a big deal but not really at the same
1:50
time you see one of the things in
1:52
programming land that we have and that
1:54
programmers love are standards and the
1:57
reason why standards are important is
1:58
they allow for us Engineers to build
2:02
systems that communicate with each other
2:04
the most popular one that you know you
2:07
might have heard of or you might not and
2:08
you don't really need to know the
2:09
details is rest rest apis and they're
2:13
basically a standard that every company
2:17
follows when they construct their apis
2:20
when they construct their services for
2:23
me as an engineer to be able to connect
The Evolution of LLMs: From Text Prediction to Tool Use
2:26
with them now understanding that
2:29
engineering is all about standards and
2:31
having these formalities we follow to
2:33
make life easier when we think of in the
2:37
context of an llm I want you to
2:39
understand this one important thing llms
2:42
by themselves are incapable of doing
2:44
anything meaningful what do I mean by
2:45
that if you remember the first you know
2:47
chat gbt 3 or was it 3.5 I'm not sure
2:51
but if you just open any chat bot and
2:53
you tell it to send you an email um it
2:56
won't know how to do that it will just
2:57
tell you hey I can't send you an email
2:59
the most you can do with an llm is ask
3:01
it questions uh maybe ask it to tell you
3:05
about some historical figure whatever it
3:08
may be right um llms are truly incapable
3:12
of doing anything uh meaningful and what
3:14
I mean by meaningful it'd be nice if you
3:16
know it could send me an email um if it
3:19
could um do some specific task on my
3:22
behalf but the only thing an llm in its
3:25
current state is good at is predicting
3:27
the next text right so for example if I
3:30
say My Big Fat Greek and llm with all
3:34
the data source with all its training
3:36
material will determine that the next
3:38
word is wedding right so this is the
3:41
most an llm by itself that it could do
3:44
right the next Evolution was developers
3:48
figured out how to take llms and combine
3:52
them with tools and you can think of a
3:55
tool like an API for example um most of
3:58
us are aware where chat gbt and these
4:00
other chat Bots are able to search the
4:02
internet for example perplexity right
4:05
perplexity gives you the option to chat
4:08
with an llm but that llm has the ability
4:11
to fetch um information from the
4:13
internet and present that to you the llm
4:16
itself is not capable of doing that but
4:19
what they've done is they've constructed
4:21
a tool they've given the llm access to
4:24
an external service right and there's
4:26
plenty of these Services right I think
4:28
there's Brave search
4:30
um chat open AI offers an API now so
4:34
llms have started to become a bit more
4:36
powerful when we connected tools to them
4:39
right I can give you an example let's
4:42
say um let's say every time I get an
4:46
email I want there to be an entry in a
4:50
spreadsheet now most of you know there
4:52
are services like zapier end8 or you
4:55
know any of those Automation
4:57
Services if I build out an autom
5:00
and connect that to my llm it just
5:03
became a bit more meaningful now that's
5:06
awesome and cool but it gets really
5:09
frustrating when you want to build an
5:11
assistant that does multiple things
5:14
imagine search the internet um read your
5:16
emails summarize this you start to
5:19
become someone who glues a bunch of
5:22
different tools to these llms and it can
5:25
get very frustrating very cumbersome if
5:28
you're wondering why we don't have have
5:30
an Iron Man level Jarvis assistant is
5:33
because combining these tools making it
5:36
work with the llm is one thing but then
5:38
stacking these tools on top of each
5:40
other making it cohesive making it work
5:43
together is a nightmare itself and this
5:46
is where we're currently at and does
5:49
before I continue does this make sense
5:50
this is where we started llms by
5:53
themselves write me a poem um you know
5:56
tell me about World War I um and then
5:59
the second evolution is oh we now have
6:01
tools right we now have um these things
6:05
these external services that we can
6:07
connect to our llm the problem here is
6:11
they're difficult it's annoying and as
6:14
someone who works at an AI startup Tempo
6:17
and we have a lot of tools like for
6:19
example we do a search um you have to
6:22
find an external service you have to
6:24
connect it to the llm and you have to
6:26
make sure the llm doesn't hallucinate or
6:28
do something stupid and believe it or
6:30
not as cool as llms are by themselves
6:34
they're very very dumb um but these
6:37
tools make them just a bit more capable
6:41
so this is where we're at uh Greg we
6:44
good so far crystal clear I'm loving
6:47
this beautiful quick break in the Pod to
6:50
tell you a little bit about startup
6:52
Empire so startup Empire is my private
6:56
membership where it's a bunch of people
6:59
like me like you who want to build out
7:03
their startup ideas now they're looking
7:05
for content to help accelerate that
7:07
they're looking for potential
7:08
co-founders they're looking for uh
7:11
tutorials from people like me to come in
7:14
and tell them how do you do email
7:16
marketing how do you build an audience
7:17
how do you go viral on Twitter all these
7:19
different things that's exactly what
7:22
startup Empire is and it's for people
7:24
who want to start a startup but are
7:26
looking for ideas or it's for people who
7:29
have a startup but just they're not
7:31
seeing the traction uh that they need so
7:34
you can check out the link to Startup
7:37
empire.co in the description now enters
MCPs explained
7:41
mCP and what does mCP mean I think the
7:45
simplest way right without getting too
7:47
technicals I've read the threads too and
7:49
as a technical person I appreciate it
7:51
but for the non- Tey I can assume it's
7:53
frustrating think of it this way think
7:56
of every tool that I have to connect to
7:58
to make my llm valuable um as a
8:02
different language so tool one's English
8:05
tool two is Spanish tool three is
8:07
Japanese right and imagine every tool
8:10
it's its own language and it's not that
8:13
there isn't a standard for how apis work
8:15
but every service provider constructs
8:18
their apis differently there's different
8:20
information you have to pass there's
8:21
just various degree of of of of things
8:25
that you have to set up that again it
8:28
just feels like gluing
8:29
a bunch of different things together
8:31
will it work yes but at scale it gets
8:34
very diff difficult mCP you can consider
8:38
it to be a layer between your llm and
8:43
the services and the tools and this
8:45
layer translates all those different
8:48
languages into a unified language that
8:51
makes complete sense to the llm right so
8:55
it's the evolution of llm plus tools but
8:58
in this Evolution it just makes it makes
9:01
it very simple for the llm to connect
9:04
and to access different outside
9:07
resources right because that's what
9:08
tools are at the end of the day so with
9:11
mCP I'm able to connect to an outside
9:14
data source an outside database maybe um
9:18
a tool like uh convex or superbase right
9:21
um imagine I I just tell the llm you
9:24
know what create me a new entry in my
9:26
database and it it's connected to my
9:30
database via mCP and it knows exactly
9:33
what to do and how to do in the second
9:36
evolution llms and tools there's a lot
9:38
of manual work that goes on there's a
9:41
lot of stepbystep planning that you have
9:43
to do and there's a lot of edge cases
9:45
where it can fail and this is why again
9:47
none of us as exciting as the space is
9:50
none of us have a Jarvis level assistant
9:52
yet it feels like we're there and we're
9:55
close but this system makes it so that
9:58
it's very diff and what's frustrating is
10:00
this imagine let me think of a simple
10:04
service a simple like you know tool
10:07
imagine um every time a slack message
10:09
comes your llm reads that slack message
10:12
and it shoots you a text right sounds
10:15
pretty trivial here's the frustrating
10:18
part imagine slack updates their API or
10:21
the text service updates makes a change
10:24
and let's say that service is connected
10:27
to other services or you have some sort
10:29
of like automation step-by-step thing
10:31
that you've
10:32
planned it becomes a nightmare it
10:35
becomes terrifying and this is why even
10:37
in the age of llms good Engineers will
10:39
still get paid because stuff like this
10:41
like this exists but what mCP does it
10:45
unifies the llm and the service right it
10:49
creates this this uh layer where the
10:54
service and the llm can communicate
10:57
efficiently now
MCP Ecosystem Overview
10:59
let's get into some practicality you can
11:02
think of the mCP ecosystem as follows
11:07
you have an mCP client you have the
11:10
protocol you have an mCP server and you
11:13
have a service right an mCP client is
11:16
something like Tempo wind surf
11:19
cursor and they are basically the client
11:24
facing side the llm facing side of this
11:27
ecosystem
11:29
the protocol again is that two-way
11:32
connection between the client and the
11:34
server and the server is what translate
11:38
translates that external service its
11:41
capabilities and what it can do to the
11:44
client and that's why between the mCP
11:46
client and the mCP server there's the
11:48
mCP protocol but here's the fascinating
11:51
part and this is why I think anthropic
11:54
they're playing 3D chess when they built
11:56
this is the way this is architected the
11:59
mCP server is now in the hands of the
12:03
service provider so if let's say me and
12:06
Greg run a Dev Tool company right where
12:09
maybe we're doing a database right like
12:11
we're like listen we're going to build
12:13
the best database company in the world
12:16
and we want people's llms to have access
12:18
to this database it is now on us to
12:22
construct this mCP server so that the
12:25
client can fully access this so
12:27
anthropic in a way sort of said listen
12:29
we want our llms to be more powerful
12:31
more capable uh but it's your job to
12:34
figure this out and this is why you've
12:36
noticed all the external service
12:37
providers are now building different mCP
12:41
servers they're building out repos and
12:42
all this stuff right so this is a big
12:46
deal in a sense where llms are going to
12:49
be more capable but from a technological
12:51
perspective all they did was create a
12:54
standard a standard that it seems like
12:56
all companies and all Engineers are
12:58
going to upon because you can construct
13:02
any system any API however you please
13:06
the problem is if you want to scale you
13:08
want to grow you want other developers
13:10
other businesses to connect and work
13:12
with your service it has to be in a
13:14
fashion that makes sense for them
13:15
imagine if all of us just spoke
13:17
different languages but standards allow
13:19
us to communicate in a way that makes
13:21
sense to all of us and mCP is that for
13:25
llms because llms by themselves are not
13:29
that capable they're just they're
13:31
they're they're they're systems that
13:33
have great predictability and they know
13:35
how to predict the next word but when
13:38
you can when you add this mCP protocol
13:41
as a whole you now have a way for it to
13:44
be capable of doing important stuff now
Technical Challenges of MCP
13:48
understanding all this it's not all
13:50
sunshine and rainbows there are some
13:52
technical challenges if you notice if
13:54
anyone has set up an mCP um server on
13:57
any of their favorite mCP clients
13:59
it's annoying um there's a lot of
14:01
downloading you have to move this file
14:03
you have to copy this that and the third
14:05
um and it's a lot of local stuff there
14:08
are some Kinks that have to be figured
14:09
out uh but once this is figured out or
14:12
finalized polished or maybe they update
14:15
the standard or maybe someone comes up
14:17
with a better one we start to enter a
14:19
world where llms start to become more
14:22
capable and that is literally all what
14:24
mCP is just making llms more capable
14:28
we're trying we're doing that with tools
14:30
right now it's kind of working um but
14:34
mCP seems to be the next Evolution I
14:36
think Greg I saw your latest video um
14:39
Manis Manis is a great example of number
14:41
two they have tons of tools and kudos to
14:46
them they've engineer engineered it well
14:49
in a way where you know they well they
14:51
work well cohesively I didn't get to try
14:53
it out so I'm just looking at what
14:55
people have done but I can tell you this
14:58
it's a lot lot of engineering hours it's
15:00
a lot of one change happens something
15:03
broke someone's on call and not sleeping
Conclusion on MCP's Potential
15:05
but with mCP um it's structured in a way
15:09
where um if we all follow this um
15:13
standard um the llm will have access to
15:16
everything it needs um and we will all
15:19
be happy users so in short that is
15:21
literally all what mCP is it's not um
15:25
Einstein's fifth law of physics or
15:27
anything crazy like that it's literally
15:29
standard for llms and it's exciting it's
15:31
something to be excited about um and
15:34
yeah I hope I hope that clarified I just
15:36
kept ramling so I apologize for that gr
15:38
no no this is this is exactly what I
15:39
wanted I want to end on one question for
15:42
you so this is now clear to me crystal
15:45
clear to me what mcps are but my
Startup Ideas for Developers and Non-Technical Users
15:49
question is well before I even ask my
15:51
question every time there's been a
15:54
popularized protocol for example https
15:58
or uh
16:00
SMTP
16:01
um examples like that there's been a lot
16:04
of big businesses that were created on
16:06
top of it and there's been basically
16:08
this like why now you know why this just
16:11
opening of opportunities yeah the
16:15
average person listening to this
16:17
podcast is building out their
16:20
ideas is this does this matter or you at
16:23
all for that person like yeah I think
16:27
that's a great question I think if I
16:29
were so I'll speak to the technical and
16:31
the non- technical to the technical
16:33
there's a lot of things that a technical
16:34
person can do here I I just don't have
16:37
time Greg but one thing I was thinking
16:38
of was like an mCP App Store and I'll
16:41
just give this idea out for free because
16:42
this this podcast is all about ideas
16:45
basically there's a lot of these repos
16:46
out there um of mCP servers and it'd be
16:50
cool if someone can go on a site I even
16:52
bought the domain um it does nothing but
16:55
again please anybody like steal this
16:58
idea um um I bought the domain and it'd
17:01
be cool if someone could um go on U like
17:06
look at the different mCP servers there
17:09
they they see the GitHub code and
17:10
whatever and they can click like install
17:12
or deploy and they that server is
17:14
deployed and gives them a specific URL
17:16
and then they can paste that in an mCP
17:18
client and work that out so for the
17:20
technical person if you make millions
17:22
all I ask is just you know send me ,000
17:25
but for the non-technical person what I
17:28
would really focus on is I would just
17:31
stay up to date with the platforms that
17:34
are building out mCP capability and just
17:37
see where the standards are going right
17:39
because like you said um when these
17:41
standards are finalized I don't know if
17:43
mCP has fully won I think it needs to be
17:46
challenged um or I don't know if
17:48
anthropic is going to make an update we
17:49
don't know it's very early but I would
17:52
say keep very close attention to what
17:55
the final standard is going to be
17:57
because once that standard is finalized
17:59
and all these service providers start to
18:03
like you know build out their mCP or
18:05
whatever thing it is you can now start
18:07
to integrate much seamlessly and much
18:10
easier right this is why again every
18:12
week There's a new chatbot interface
18:14
with new tools and it wins because this
18:17
part step number two is not easy right
18:20
especially making it cohesive and making
18:22
it work fast right like I can sit in two
18:25
hours and build something like this but
18:28
building out that user experience making
18:30
it Flawless limiting the hallucinations
18:33
it's very very hard I mean this is a lot
18:36
of the work we do at Tempo but this
18:40
makes it so that integrating is a lot
18:43
easier and you can think of these as
18:46
like Lego pieces that you can continue
18:49
to stack to stack so for my smart and
18:52
wise business owners startup uh ideas
18:56
podcast enjoyers I would really just
18:58
keep a close attention right I think
19:01
even for myself I don't think with this
19:03
mCP stuff we're at a place where any
19:06
shots can be fired that make um any
19:09
smart business decision but this is one
19:11
of those things where you just you sit
19:13
and you watch and you're just observing
19:15
and learning and when the right thing at
19:17
the right time happens you strike so I
19:19
don't see any crazy business
19:21
opportunities right now for a
19:23
non-technical person even for a
19:25
technical person like imagine if open AI
19:27
comes with a standard tomorrow and we
19:28
all just shift to that right it's very
19:31
early stages but I think understanding
19:33
how this works means you'll understand
19:35
how the next thing works and when that
19:37
becomes fin finalized you hit the ground
19:40
running amen all
19:42
right Ross Mike Professor Ross Mike
19:46
there's no one like you we'll include in
19:48
the show notes where you can follow him
19:50
for more really clear explanations
19:53
around this whole AI coding world and uh
19:58
dude I'll see you in Miami in a few
20:00
weeks yeah man I appreciate you I'm
20:02
booking my flight soon so yeah
20:03
definitely bro I'll see you soon thank
20:05
you everybody
20:07
[Music]
view_moduleinsert_chart