MCP Learning Series - Part 4

-

Now that we have an idea on what is MCP and how to create a simple MCP server and add it to AI Application like Claude.

But, Do we really need a AI Application to access MCP Server?

Answer is NO

AI Applications like Claude are using LLM for thinking, reasoning and acting. 

MCP is a Client - Server architecture.

Where the client is your AI application. But, to be precise these AI Applications are called "MCP Host".

These MCP host contains inbuild MCP client.


The above diagram explains the architecture of MCP.

1) After we add the weather MCP server to the AI Application - MCP Host.

2) MCP Client creates a dedicated connection (1:1) with the weather MCP Server.

3) This connection exist, regardless of the MCP Server is used or not.

4) When the user enters the prompt, "What is the weather in Chennai?"

5) The LLM discovers the tools available and based on the input context and tool description. It uses the weather tool.

6) This call is forwarded from the MCP Host (LLM) -> MCP Client -> MCP Server (Weather)

7) The MCP Server responds back to the MCP Client.

8) The MCP Client forwards to LLM.

9) LLM responds back to the user.


Key points:

The LLM never talks to the server directly.

The LLM chooses the tool from the descriptions.

There can be multiple rounds - Model may call the tool twice before answering.

No tool is called if it isn’t needed.

The connection is typically already open.

Sample MCP Client: To access the MCP Server without AI Application

import asyncio
import sys

from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client


async def main():
    params = StdioServerParameters(command=sys.executable, args=["main.py"])

    async with stdio_client(params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()

            #result = await session.list_tools()
            #for tool in result.tools:
            #    print(tool.name, "-", tool.description)
            #    print("  schema:", tool.inputSchema)
            result = await session.call_tool(
                "search_news",
                {"query": "Anthropic", "max_results": 3, "lang": "en"},
            )

            for item in result.content:
                if item.type == "text":
                    print(item.text)


asyncio.run(main())


Comments

Post a Comment

Popular posts from this blog

Agentic AI - Series 3

-
Image
                                                                      In this blog we will see how to build a simple agent without using any LLM model. Simple flow of an agent: Agent takes input -> Based on the input, it decides which tool to use -> Performs operation and send the output back to the user. An AI agent framework is  a set of tools, libraries, and structures that simplifies building, deploying, and managing autonomous AI agents. But, here we are going to build an agent without using any framework and LLM's. Agent functionality is to perform "Addition" and "Subtraction". Agent uses 2 python functions (Tools) to perform addition and subtraction operations. Finally, it send the output to the user. We are going to implement the above discussed functionality vi...
Read more

Agentic AI - Series 4

-
Image
                                                                   In this blog, will how to create a simple agent using AI framework like Langchain with a simple LLM model  gpt-4o-mini. from langchain_openai import ChatOpenAI from langchain_core . tools import tool import os # Creating a tool @ tool def add_numbers ( x : int , y : int ) -> int :     "Add two numbers"     return x + y @tool is a decorator to define a tool. This is a simple tool to add 2 numbers. The below line creates a Langchain LLM wrapper around on the OpenAI chat mode 'gpt-4o-mini' # Binding the tool to a model llm = ChatOpenAI( model = "gpt-4o-mini" ).bind_tools([add_numbers]) response = llm .invoke( "What is 5 + 7?" ) When we print response, is the expected outcome? Is it 12. NO...
Read more

Principal SRE - Interview Question

-
Image
  Sharing some of the interview questions for the role of Principal SRE from Apple: 1) Tell me about a production outage you’ve handled that had ambiguous symptoms. How did you narrow it down? 2) How do you decide what belongs in an SLO, and how do you avoid overengineering it? 3) You inherit a platform with strong uptime but high operational toil. What do you change first? 4) How do you handle a disagreement with product leadership when reliability work competes with feature delivery? 5) A service is scaling rapidly and latency is degrading under load. Walk me through your approach. 6) What does “good observability” mean to you in practice? 7) Describe how you would lead a major incident. 8) What’s your philosophy on automation in production operations? 9) How do you evaluate whether an architecture is resilient enough? 10) You’re the on-call SRE for a globally used service. At 2:00 AM, error rates jump from 0.2% to 8%, latency doubles, and one region is still healthy wh...
Read more