I keep hearing about MCP and how useful

it is for AI apps. At this point, I'm

almost afraid to ask, what is it and how

does it work?

>> Well, Martin, let me show you.

[Music]

>> Welcome to the show, Jack. What do you

do here at Google?

>> Thanks for having me, Martin. Well, I'm

a Python devel developer advocate here

at Google Cloud and my background is in

open source client libraries

specifically on databases and making the

connection experience smooth.

>> Excellent. So, what is MCP and why is it

useful?

>> Yeah. Well, MCP stands for model context

protocol. It's an open protocol

developed by Enthropic that standardizes

how we provide context to large language

models or LLMs. There are three main

components to MCP, but today we are

going to focus on tools.

>> Okay, tools. That sounds a bit abstract

to me. What What are they? Well, tools

give an LLM or agent the ability to take

action in the real world through

discovering and calling external APIs,

querying a database, or running custom

code, you name it.

>> That makes sense. Who would use MCP?

MCP really is useful for a wide

audience, but I see most users falling

into one of two groups of people. If you

provide an API or a service, you may

want to use MCP to package your API into

MCP tools so that clients and agents

have an easier time using it. If you are

a developer using AI assisted or

building AI agents, MCP makes it easier

for agents to use an API or perform

actions on your behalf. For example, an

MCP server for Cloud Run. A developer

can ask in plain English to deploy a

Cloud Run service instead of having to

know all those pesky G-Cloud commands.

>> Those pesky G-Cloud commands. Yes, I

feel you. Uh, and where would you deploy

this MCP server?

>> Good question. You have lots of

different options. You can run an MCP

server locally. For example, if you want

AI help when you are writing code in

your editor. You can also have remote

MCP servers if you want to share it with

others. I think deploying MCP to Cloud

Run on Google Cloud is a great option.

>> All right, let's get down to brass tax.

What might a real use case of MCP look

like?

>> Yeah, let's say Alice is an API

provider. Her company sells sprinkler

systems for lawns. She partners with

lawn care companies who install these

sprinkler systems. To make life easier

for them, she provides an MCP server

with a tool that they can call to get a

price quote. She has deployed the MCP

server on Cloud Run, which is

serverless, so that she doesn't have to

worry about the infrastructure.

>> Got it.

>> David runs a lawn care company. He

provides an AI chatbot that his

customers can use to get lawn care tips,

to schedule visits, or to place orders,

and so on. If David points his AI agent

to Alice's MCP server, the agent can now

provide quotes for installing sprinkler

systems.

>> Oh, shadow, please get down. Okay. All

right. Sorry about that, Jack. I see.

Uh, so what would the end user

experience be?

>> Yeah. Let's say we have Eve, who is the

end user. She might use David's AI

chatbot to ask for general lawn care

tips, like when is the best time to

water her lawn.

She might also ask how much it would

cost to install a new sprinkler system.

David's chatbot can actually answer this

question thanks to Alice's MCP server

and API. That makes sense. Now, I'm a

developer. What does the code for all

this look like?

>> That's an important part. Let's start

with Alice's MCP server. She has

imported the fast MCP library and added

this MCP.tool decorator to the calculate

quote function. That annotation is all

it takes to expose this method via the

MCP protocol. So AI agents can now use

it. The body of that function is the

business logic for calculating the quote

itself. The price quote depends on which

country the customer in as well as the

square meters their lawn.

>> All right. Uh, that looked pretty

simple. Now, how would David's agent

call Alice's tool to calculate a quote?

>> Let's take a look at the actual code for

David's agent. When you build an agent,

you can provide an array of tools or

functions that it can call. David's

provides a single tool of type MCP tool

set which contains the address of

Alice's MCP server.

>> Okay. And when I've given tools to

agents before, I remember having to

provide which input and output

parameters each function takes and a

description of what each function does.

Uh, but David didn't have to do that

here.

>> That's right. He didn't. He just pointed

the agent to Alice's MCP server URL and

it figured out the rest.

>> And then David's agent called Alice's

API.

>> Yep, it did. Alice is running her MCP

server on Cloud Run. So, if she clicks

on the Cloud Run service logs in the

console, here's what it looks like. Here

is the actual call made by David's

agent.

>> That makes sense. And where would David

run his agent?

>> Another good question. The agent was

built using Google's ADK or agent

development kit and it was also deployed

to Cloud Run.

>> Wow, a lot of Cloud Run in this example,

huh?

Yeah, Cloud Run is serverless which

makes it really easy to deploy agents,

MCP servers or APIs. Google handles the

configuration and scaling so that you

can release faster.

>> Very good. But I have some questions for

you Jack.

>> All right. All right. Go ahead, Martin.

>> So, the point of MCP is that agent

developers like David don't have to tell

the AI what the input and output

parameters are.

That's one advantage of MCP. Another

reason for MCP is that previously your

tools had to be built for a specific

framework like lang chain or visual

studio code. MCP is a common protocol

that many clients can use. On the other

side, developers who expose their APIs

with MCP. That way, Alice's API or the

API for a calendar, email, or local file

system all appear the same and make it

easier for AI agents. M and you said

before that we'd put the MCP server in

the cloud to make it available for

everyone, but Alice might want some

authentication for MCP server, so only

registered partners like David can use

it right?

Yes. Luckily, MCP has support for

authentication. I will note that it is

fairly new and still evolving. MCP

supports OOTH, which is great for when

users like you or me want to

authenticate. You can also lean into

cloud run existing authentication infra

to use with MCP like bear off. This is a

great choice when you are wanting

servicetoervice authentication.

>> I see. What does this MCP protocol

actually look like?

>> Good question. Let's take a look. I have

the MCP inspector pulled up on my

computer. Here we can see that the tools

is an array with a single entry called

calculate quote. That tool takes two

inputs, square meters and country code.

Both are required.

>> That looks a little bit like an open API

spec. What's the difference?

>> That's an important question. An open

API spec is lower level. For example, it

includes the HTTP status codes that an

API can return. The MCP server skips

those implementation details.

>> In the code you showed, Alice used uh

Python and the fast MCP library. All is

good. But what if Alice already has a

well- tested API and doesn't want to

change it? Uh, she just wants to expose

her existing API to AI agents.

>> Totally. Fast MCP actually has built-in

support for this exact functionality.

You can point it to your open API spec

and it will actually generate an MCP

server. Another option is to write a

small proxy using fast MCP that forwards

any calls to the underlying API.

>> Very good. Now, let's recap. What are

the main takeaways here, Jack?

>> As I was first learning about MCP and

writing my first MCP servers, here's

what I learned along the way. One, MCP

is super useful as a common language for

declaring tools for our AI agents. Two,

if you want agents to use your API,

expose it with MCP. And three, it's

super easy to run an MCP server on Cloud

Run.

>> Great takeaways and thanks for joining

me today, Jack. Thanks for having me,

Martin.

>> And thank you everyone for watching. If

you have any questions for Jack or me,

ask in the comments. And do let me know

what you thought of today's episode. I

read every single comment. I can't wait

to see what you build.

[Music]