Building a Tool-Integrated AI Assistant with Model Context Protocol (MCP) and Google Gemini 2.5 Pro

As Large Language Models (LLMs) like GPT-4, Claude, Gemini, and Llama 3 continue to advance, the need for a standardized way to connect them to real-time data sources, APIs, and tools becomes essential. These models, while powerful, are isolated—they don't naturally interface with the outside world.

In a previous post, I introduced Function Calling as a way to bridge this gap. Today, we'll take that further with the Model Context Protocol (MCP) — a robust, open protocol that transforms how LLMs interact with external systems.

What Is Model Context Protocol (MCP)?

Developed by Anthropic, MCP acts like a "USB-C for AI tools", offering a universal interface for connecting LLMs to external APIs, databases, files, and services.

MCP Architecture:

• Clients: LLMs or AI applications
• Servers: Host tools/APIs and expose them to clients

This client-server design enables structured, dynamic tool interactions across systems.

Key Benefits:

• Standardization: Unified method for connecting LLMs to APIs
• Modularity: Reuse and compose tools with ease
• Security: Clean handling of API keys and sensitive data
• Interoperability: Works across languages and frameworks
• Discoverability: Tools self-describe via schema, so no more guesswork

When Should You Use MCP?

• Building AI agents or assistant frameworks
• Integrating with multiple tools (e.g., search, databases, file systems)
• Scaling systems with reusable and discoverable tools
• Replacing hardcoded integration logic with dynamic function execution

Real-World Example: Flight Search with Gemini + MCP

Let's walk through building a natural language flight search tool powered by Gemini 2.5 Pro and an MCP server.

Tools:

• Gemini API – for language understanding + function calling
• SerpAPI – to get real-time flight data
• MCP Server (`mcp-flight-search`) – a FastMCP server wrapper

Step-by-Step Implementation

✅ Prerequisites:

• Python 3.8+
• Google Gemini API key
• SerpAPI key

python -m venv venv
source venv/bin/activate
pip install google-genai mcp mcp-flight-search

from google import genai
from google.genai import types
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

client = genai.Client(api_key=os.getenv("GEMINI_API_KEY"))

Best Practices for Using MCP + Gemini

🛠 Tool Design

• Use descriptive names (`search_flights`)
• Keep input schema simple and well-typed
• Make descriptions model-friendly

🔄 Model Usage

• Don't overload with too many tools
• Dynamically load tools based on context
• Prompt Gemini with purpose (e.g., "You are a travel assistant...")

⚙️ Server & Security

• Use `stdio` for ease of development
• Pass API keys using env vars
• Log clearly but securely

Limitations (as of March 2025)

• Partial OpenAPI support
• Function calling only available in Python SDK
• Some parameter types not yet supported

GitHub Repository

All source code and the working demo can be found here:
👉 GitHub – arjunprabhulal/mcp-gemini-search

Conclusion

By combining Google Gemini with the Model Context Protocol, we created an intelligent AI agent that:

• Understands natural language
• Dynamically discovers tools
• Fills in parameters via function calling
• Executes real-world actions in real time

This pattern enables you to build extensible, secure, and modular AI systems that go beyond chat — into automation, reasoning, and real-world utility.