Model Context Protocol | Build MCP Server | Integrate MCP server with AI Agents - Local and Remote | Sagar Kakkala´s World
.png)
Model Context Protocol
Before we discuss about MCP, let us discuss the use case of it, lets say you have chatbot or an AI Agent, imagine you giving a prompt like "Play best horror film in Netflix" or "Book a flight from Delhi to Hyderabad under 3000 rupees". Imagine your chatbot searching flights, comparing flight prices and booking tickets for you and making you avoid lot of trouble.
And this is possible when your AI Agent has MCP Tool attached to it. MCP can either be local or even remote which we will understand it better later in this demo
lets say you are running an Airlines or any hotel, May be any business that you are running, You can design your own MCP tool and publish it to use. And third party can integrate your MCP tool and make efficient use of your MCP tool which can help in your business profits as well.
MCP Components
MCP stands for Model Context Protocol, Model can be any LLM that you specify or agent can use model of its own choice. Context is telling what your model is designed to do and Protocol are some rules that are meant to be followed.
To design and MCP server, You need to have Resources, Tools , Prompts
We will understand this better when we implement the project.
let us say, we are running a hotel booking app business like Airbnb, Booking.com... and resources are nothing but data that we have, it can be details of particular hotel in particular place and prices of hotel.
let us take sample data for example
The Data of Hotels can be resources
Hotels
Grand Palace Hotel – New York
• Standard: $200
• Deluxe: $350
• Suite: $500-
Ocean View Resort – Miami
• Standard: $180
• Deluxe: $300
• Suite: $450 -
Mountain Lodge – Denver
• Standard: $150
• Deluxe: $250
• Suite: $400 -
City Central Inn – New York
• Standard: $170
• Deluxe: $280
• Suite: $420 -
Sunset Boulevard Hotel – Los Angeles
• Standard: $190
• Deluxe: $320
• Suite: $480
Now based on above data, we can design few tools like , A tool that can help get hotel details, another tool that can get prices of hotel, and also tool to book Hotel.
Depending on business and use case, you can design your own tools using MCP
and now we can also use prompts, to make our MCP server aware that user can search the following way and to be ready for it, it is more like expected input for the MCP to work
like user may ask "Please help me find the best hotel in {city} within a ${budget} budget" and what action to be taken is actually what we pre-design in prompt
Build Local MCP Server
let us understand it more clear by building MCP server using MCP Python SDK
Note: If you want to Build MCP server using different SDK, you can checkout official documentation - MCP SDK
Install Jupyter Notebook on your local using Anaconda if you dont have it installed already- Anaconda
We will be using Jupyter Notebook, throughout our project and we will be installing uv for this particular project
UV is actually written in RUST and modern package alternate for PIP, which is reliable and fast
For MacOS
curl -LsSf https://astral.sh/uv/install.sh | shFor Windows
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"Let us use terminal in Jupyter Notebook as we will need terminal throughout our demo
click on New > Terminal
Once Terminal is opened, copy our UV install command as given in above code snippet
since i had UV installed already, you might have different output log
Now let us go for our MCP Python SDK readme file , you will find MCP installation commands
let us modify commands a bit for our hotel booking app
you can either create a folder , navigate to it and run "uv init" command in it or you can use something like "uv init hotel-booking", for uv to create folder directly
We will be creating inside AI Agents folder that we created in our last session - AI Agents session
Now we will be running commands inside our Terminal
Navigate to directory where you want to create MCP server
and now we follow commands given in readme with slight modifications required for our local MCP sever
uv init hotel-booking
this creates a hotel-booking folder and you can verify the files inside it via Jupyter Notebook and you will be able to observe python files created inside directory
Now let us get back into our terminal to make this project ready
Navigate to freshly created directory and install mcpcli
cd hotel-bookinguv add "mcp[cli]"
and now you will be able to see uv.lock file once the command is run and it has all dependencies required
and now you can use MCP server that is mentioned in readme file for test, let us build our own MCP sever
create a new file "server.py" or any name of your choice
and now copy paste contents below in server.py - this is our main file that serves as our MCP server which has details required
from mcp.server.fastmcp import FastMCP
# Initialize FastMCP app
mcp = FastMCP("Hotel Booking Server")
# -------------------- RESOURCES --------------------
HOTELS = [
{"id": 1, "name": "Grand Palace Hotel", "city": "New York", "rating": 4.5},
{"id": 2, "name": "Ocean View Resort", "city": "Miami", "rating": 4.7},
{"id": 3, "name": "Mountain Lodge", "city": "Denver", "rating": 4.2},
{"id": 4, "name": "City Central Inn", "city": "New York", "rating": 4.0},
{"id": 5, "name": "Sunset Boulevard Hotel", "city": "Los Angeles", "rating": 4.3}
]
PRICES = {
1: {"standard": 200, "deluxe": 350, "suite": 500},
2: {"standard": 180, "deluxe": 300, "suite": 450},
3: {"standard": 150, "deluxe": 250, "suite": 400},
4: {"standard": 170, "deluxe": 280, "suite": 420},
5: {"standard": 190, "deluxe": 320, "suite": 480}
}
@mcp.resource("data://hotel_details")
def hotel_details():
"""Return all hotel details"""
return HOTELS
@mcp.resource("data://hotel_prices")
def hotel_prices():
"""Return all hotel price details"""
return PRICES
# -------------------- TOOLS --------------------
@mcp.tool()
def get_hotels(city: str) -> dict:
"""Get hotels available in a city"""
city_lower = city.lower()
results = [h for h in HOTELS if h["city"].lower() == city_lower]
if not results:
return {"message": f"No hotels found in {city}"}
return {"hotels": results}
@mcp.tool()
def get_prices(hotel_id: int) -> dict:
"""Get room types and prices for a hotel"""
data = PRICES.get(hotel_id)
if not data:
return {"error": f"No price information for hotel_id {hotel_id}"}
return {"hotel_id": hotel_id, "prices": data}
@mcp.tool()
def book_hotel(hotel_id: int, guest_name: str) -> dict:
"""Book a hotel room for a guest"""
hotel = next((h for h in HOTELS if h["id"] == hotel_id), None)
if not hotel:
return {"error": "Hotel not found."}
confirmation = f"CONF-{hotel_id}-{guest_name[:3].upper()}"
return {
"message": "Booking successful!",
"hotel": hotel["name"],
"guest": guest_name,
"confirmation_id": confirmation
}
# -------------------- PROMPTS --------------------
@mcp.prompt()
def find_best_hotel(budget: float, city: str, preferences: str = "standard") -> str:
"""Generate a prompt for finding the best hotel within budget"""
return f"""Please help me find the best hotel in {city} within a ${budget} budget.
Preferences: {preferences}
Consider:
- Price (must be under ${budget})
- Hotel rating and reviews
- Room type and amenities
- Location convenience
Use the get_hotels and get_prices tools to find options and provide a recommendation with reasoning."""
@mcp.prompt()
def handle_booking_issue(hotel_name: str, issue: str) -> str:
"""Generate a prompt for handling hotel booking issues"""
return f"""A guest's booking at {hotel_name} has an issue: {issue}.
Please help resolve this by:
1. Understanding the guest's situation
2. Offering alternative hotels or room options
3. Providing clear rebooking or refund steps
4. Ensuring the guest feels accommodated
Be empathetic and solution-focused in your response."""
# -------------------- RUN MCP --------------------
if __name__ == "__main__":
print("🏨 Starting Hotel Booking MCP Tool...")
mcp.run()
As from above file, you can verify that there are decorators namely @mcp.resources , @mcp.tools, @mcp.prompts which stands as important for our MCP server and make sure you save the file.
Now to test if our MCP server works fine, we can use MCP inspector and before we use MCP inspector, please do install nodejs and npm from terminal
For MacOS
brew install nodeFor Windows
choco install nodejs
Now to use MCP inspector, run the following command in terminal
uv run mcp dev server.py
Now this would open MCP inspector on your browser
click on connect
once connected, you can see Tools, Resources, Prompts and other configurations
click on List Resources
similarly you can verify Prompts and Tools
This shows that we have built our MCP server successfully
Integrate Local MCP tool with AI Agent
Now before we move further into this demo, understand how to configure AI Agents, How to use Handoff Agents, How to Integrate AI Agents with tools - AI Agents
Now here in this demo, i will be using same setup that we had for AI Agents configured, which mean open-ai API key is configured, Prerequisites required for AI Agent has been done already
and since we are doing local MCP setup with open-ai agents, you can refer to official documentation here - openai agents python mcp
since ours is local MCP setup, we will be using stdio MCP servers referring to the documentation
Now since we are using uv for this project to run our MCP project, we will make modifications accordingly.
Also know exact path of your MCP server locally , here path i would refer is /Users/sagar/AI Agents/hotel-booking
and now let us implement our agent, if you have not configured AI Agent already, make sure you get OpenAI API key, save it an .env file as showed in previous project
and run following commands
from dotenv import load_dotenv
load_dotenv()and install open-ai agents
pip install openai-agentsAI agent tool with local MCP
from pathlib import Path
from agents import Agent, Runner
from agents.mcp import MCPServerStdio
import asyncio
current_dir = Path("/Users/sagar/AI Agents/hotel-booking") # path to server.py
chat_history = []
async def chat_with_hotel_agent():
# Start your local MCP server in stdio mode
async with MCPServerStdio(
name="Hotel MCP Server",
params={
"command": "uv",
"args": ["run", "python", "server.py"], # command to start your MCP server
"cwd": str(current_dir),
},
) as server:
# Define the AI agent using this MCP server
hotel_agent = Agent(
name="Hotel Booking Agent",
instructions="You help users find hotels, check prices, and make bookings using the MCP Hotel tool.",
mcp_servers=[server],
)
print("🏨 Enter the hotel query. Type 'exit' to quit or 'history' to see previous chats.\n")
while True:
user_input = input("You: ")
if user_input.lower() in ["exit", "quit"]:
print("Goodbye! 👋")
break
elif user_input.lower() == "history":
print("\n---- Chat History ----")
for i, (u, b) in enumerate(chat_history):
print(f"{i+1}. You: {u}\n Agent: {b}")
print("----------------------\n")
continue
# Send the message to the hotel agent
result = await Runner.run(hotel_agent, input=user_input)
response = result.final_output
# Save to chat history
chat_history.append((user_input, response))
print(f"Agent: {response}\n")
# In Jupyter, call with await
await chat_with_hotel_agent()
Now after running this AI Agent, it opens chatbot and logic behind code has been explained in AI Agents session and we will be concentrating on integrating MCP servers in this Demo
As since we have integrated MCP tool, it must be able to show us hotel data, hotel prices from our local itself
let us verify now
As you can see from Below screenshot, it was able to book hotel for me and it gave me confirmation ID as well
it pulled up data from resources we provided and used tool that we specified.
AI agent with Local MCP sever to use only particular tool
Lets say, in this particular project, we had three tools for AI Agent to use and consider MCP server with lot of tools and giving access to AI Agents to lot of tools can raise security concerns if customer decide to use it with wrong intentions
You may configure as many tools in MCP but you can also restrict tools used here with static filters or even block particular tools
let us improvise our code accordingly to use only get_hotels tool which means it cant book hotel or check prices of hotel
from pathlib import Path
from agents import Agent, Runner
from agents.mcp import MCPServerStdio, create_static_tool_filter
import asyncio
current_dir = Path("/Users/sagar/AI Agents/hotel-booking")
chat_history = []
async def chat_with_hotel_agent_filtered():
# Start MCP server in stdio mode with tool filtering
async with MCPServerStdio(
name="Hotel MCP Server",
params={
"command": "uv",
"args": ["run", "python", "server.py"],
"cwd": str(current_dir),
},
tool_filter=create_static_tool_filter(allowed_tool_names=["get_hotels"])
) as server:
hotel_agent = Agent(
name="Hotel Finder Agent",
instructions="You help users find hotels in a city using the MCP get_hotels tool.",
mcp_servers=[server],
)
print("🏨 Enter the city to find hotels. Type 'exit' to quit.\n")
while True:
user_input = input("You: ")
if user_input.lower() in ["exit", "quit"]:
print("Goodbye! 👋")
break
result = await Runner.run(hotel_agent, input=user_input)
response = result.final_output
chat_history.append((user_input, response))
print(f"Agent: {response}\n")
await chat_with_hotel_agent_filtered()
Now let us verify if it can book hotels now for us.
As you can see from above screenshot, Our Agent was able to get hotel details but not able to book hotel ,it is by default we only allowed one tool that is to get hotels.
Integrate AI Agent with Remote Server
since this is local MCP server, we can also use MCP server configured by that particular business, you can check list of MCP servers available here in readme file - Remote MCP Servers
and now since there is chrome dev MCP server, let us integrate that MCP tool
you may choose to integrate any other MCP tool that you are interested in, here i will be using Chrome Dev Tools - Chrome Dev MCP
When you scroll down to readme file, you will be able to find how to integrate Chrome MCP server
Now let us make modifications and use this MCP tool to check Airbnb available in particular location
from pathlib import Path
from agents import Agent, Runner
from agents.mcp import MCPServerStdio
import asyncio
chat_history = []
cwd = Path("/Users/sagar/AI Agents/hotel-booking") # folder where you want the server to run
async def chat_with_airbnb_agent():
# Start Airbnb MCP server via npx
async with MCPServerStdio(
name="chrome MCP Server",
params={
"command": "npx",
"args": ["-y", "chrome-devtools-mcp@latest"],
"cwd": str(cwd),
}
) as server:
# Define the AI agent using this MCP server
airbnb_agent = Agent(
name="airbnb Agent",
instructions="You help users search Airbnb listings, check prices, and make bookings using the chrome search tool.",
mcp_servers=[server]
)
print("🏠 Ask me to search Airbnb in chrome listings. Type 'exit' to quit.\n")
while True:
user_input = input("You: ")
if user_input.lower() in ["exit", "quit"]:
print("Goodbye! 👋")
break
result = await Runner.run(airbnb_agent, input=user_input)
response = result.final_output
chat_history.append((user_input, response))
print(f"Agent: {response}\n")
# Run in Jupyter
await chat_with_airbnb_agent()
Now we have asked our AI Agent to use chrome tool and search for Airbnb, also make sure you chrome installed in your laptop for this AI Agent to use chrome,
I gave prompt to search for Airbnb in Bengaluru and it opened chrome tool to show me
You can experiment using chrome tool the way you want it to work, let us use same MCP tool for different purpose, you can use the following codes to use it accordingly
Youtube Agent that can help you search for paticular videos
from pathlib import Path
from agents import Agent, Runner
from agents.mcp import MCPServerStdio
import asyncio
chat_history = []
cwd = Path("/Users/sagar/AI Agents/hotel-booking") # folder where you want the server to run
async def chat_with_youtube_agent():
async with MCPServerStdio(
name="chrome MCP Server",
params={
"command": "npx",
"args": ["-y", "chrome-devtools-mcp@latest"],
"cwd": str(cwd),
}
) as server:
# Define the AI agent using this MCP server
youtube_agent = Agent(
name="youtube Agent",
instructions="You help users play videos on youtube that they have asked you and you use chrome tool for it.",
mcp_servers=[server]
)
print("What do you love to watch on Youtube today? . Type 'exit' to quit.\n")
while True:
user_input = input("You: ")
if user_input.lower() in ["exit", "quit"]:
print("Goodbye! 👋")
break
result = await Runner.run(youtube_agent, input=user_input)
response = result.final_output
chat_history.append((user_input, response))
print(f"Agent: {response}\n")
# Run in Jupyter
await chat_with_youtube_agent()
Netflix Agent that can help you find particular film in Netflix
from pathlib import Path
from agents import Agent, Runner
from agents.mcp import MCPServerStdio
import asyncio
chat_history = []
cwd = Path("/Users/sagar/AI Agents/hotel-booking") # folder where you want the server to run
async def chat_with_netflix_agent():
async with MCPServerStdio(
name="chrome MCP Server",
params={
"command": "npx",
"args": ["-y", "chrome-devtools-mcp@latest"],
"cwd": str(cwd),
}
) as server:
# Define the AI agent using this MCP server
netflix_agent = Agent(
name="netflix Agent",
instructions="You help users search for videos on Netflix that they have asked you and you use chrome tool for it.",
mcp_servers=[server]
)
print("What do you want me to search on Netflix today? . Type 'exit' to quit.\n")
while True:
user_input = input("You: ")
if user_input.lower() in ["exit", "quit"]:
print("Goodbye! 👋")
break
result = await Runner.run(netflix_agent, input=user_input)
response = result.final_output
chat_history.append((user_input, response))
print(f"Agent: {response}\n")
# Run in Jupyter
await chat_with_netflix_agent()
similarly you can design your own Spotify search agent and to have no limitation using this MCP tool to your agent
This concludes the blog
📚You can now Follow me on LinkedIn for latest updates
📜You can now Follow me on Instagram for behind the scenes
📹 Subscribe my Youtube Channel to stay up to date with contents
Comments
Post a Comment