Practical Usage of
AI Agents and MCP Protocol

From Theory to Working Code in 30 Minutes

Muhammad Abdugafarov
October 11, 2025
Khujand, Tajikistan

Scan to view slides

About Me 👨‍💻

Name: Muhammad Abdugafarov

Role: Team Lead at SilkRoad Professionals

Experience: 8 years in IT

Interests: AI, product development

What is an LLM? 🧠

Large Language Model

Neural network trained on massive amounts of text
Understands and generates natural language text
Examples: GPT-4, Claude, Llama, Gemini
Can perform various tasks through text interface

💡 LLM = "Brain" that understands human language

What Can LLMs Do? ✨

Can:

Understand text, images, audio
Generate responses
Analyze data
Write code
Translate languages

Cannot:

Perform real-world actions
Use external tools
Change system state
Work with real-time data

💡 LLMs can think, but can't act

From LLM to AI Agent 🤖

LLM (Just a model)
    ↓
    + Tools
    + Memory
    + Decision-making ability
    ↓
AI Agent (Can act)

AI Agent = LLM + ability to interact with the world

What is an AI Agent? 🤖

Components:

🧠 LLM (brain)
🛠️ Tools (hands)
💾 Memory (context)
🎯 Goal (task)

Capabilities:

Performs actions
Makes decisions
Uses tools
Solves complex tasks

Agent = Not just a chatbot, but an assistant that can act!

The Problem 🤔

LLMs are inherently limited

Limited by training data (knowledge becomes outdated)
Cannot interact with external world
No access to your systems and data
Cannot perform real actions

Question: How to give LLMs the ability to use tools?

Solution: MCP ✨

Model Context Protocol

Standard protocol for connecting tools to LLMs
Gives LLMs the ability to interact with external world
Created by Anthropic, works with any LLM
Transforms LLMs into AI agents

MCP = USB for AI tools

What is MCP? 🔌

Model Context Protocol
Standard for AI-to-tool communication
Like REST API, but for AI agents
Server provides tools
AI automatically discovers and uses them

System Architecture 🏗️

Why MCP? 💡

With MCP ✓

Standardized protocol
Plug-and-play tools
Works with any LLM
Easily extensible

Without MCP ✗

Custom integration
For each tool
Tightly coupled to LLM
Hard to maintain

Today's Demo 🎬

Expense Tracker Agent

Track expenses in natural language
"I spent 50 on groceries" → Creates transaction
AI automatically categorizes
Provides financial insights
No forms, no clicks - just conversation

System Architecture 🔧

User → AI Agent (Agno) → MCP Server → FastAPI → SQLite
                ↓
          OpenAI GPT-4
                ↓
        Processes request
        Selects tools
        Formats response

Tech Stack 🛠️

Agent Framework: Agno
AI Model: OpenAI GPT-4
Protocol: Model Context Protocol (MCP)
Backend: FastAPI + SQLite
Clients: Web UI, Telegram Bot
Language: Python 🐍

MCP Server - code 💻

from fastapi import FastAPI
from fastapi_mcp import FastApiMCP

app = FastAPI()

# That's it! 5 lines of code
mcp = FastApiMCP(
    app,
    name="Expense Tracker MCP"
)

# Automatically converts all FastAPI routes into MCP tools!

FastAPI Backend 🚀

@app.post("/transactions")
async def create_transaction(
    data: TransactionCreate
):
    return store.create(data)

# Regular FastAPI route
# Becomes MCP tool automatically
# AI can call it via MCP

Agent Configuration ⚙️

agent = Agent(
    name="Expense Tracker Agent",
    model=OpenAIChat(id="gpt-4"),
    
    # Connect to MCP server
    tools=[MCPTools(
        url="http://localhost:9002/mcp"
    )],
    
    # System instructions
    instructions=SYSTEM_PROMPT
)

🎬

LIVE DEMO

Let's see how it works!

Under the Hood 🔍

1. User: "Show my balance"
   ↓
2. GPT-4 thinks:
   - Need tool: get_summary
   ↓
3. MCP call:
   POST /mcp/tools/get_summary
   ↓
4. FastAPI executes:
   SELECT SUM(...) FROM transactions
   ↓
5. Agent responds:
   "Your balance: 5000.00"

Real-World Use Cases 🌍

Customer Support: AI agent with CRM tools
DevOps: AI with monitoring and deployment tools
Data Analysis: AI with databases and analytics
Content Management: AI with CMS tools
IoT Control: AI with device management tools

Production Benefits 📈

Reduced integration time (hours vs weeks)
Consistent tool interface
Easy to add new capabilities
Model-agnostic (easily switch GPT to Claude)
Better AI tool selection

Try It Yourself! 🚀

# Clone the repository
git clone https://github.com/dev-muhammad/MCPAgent
cd MCPAgent

# Configure
cp .env.example .env
# Add OPENAI_API_KEY to .env

# Run
cd server && python start.py --mcp
cd server && python start.py --api
cd agent && python agent.py

Resources 📚

Documentation:

MCP: modelcontextprotocol.io
FastAPI-MCP: github.com/jlowin/fastapi-mcp
Agno: docs.agno.com

GitHub Repository:

github.com/dev-muhammad/MCPAgent

Questions? 🤔

Thank you for your attention! 🎉

Practical Usage ofAI Agents and MCP Protocol