Deploying an Agentic LLM with MCP and LangGraph

This guide walks through how to use LangGraph with LangChain MCP to deploy a tool-using agentic LLM. We’ll use a local model via Ollama, wrap tools using MCP, and connect everything into a reactive LangGraph agent.

🚀 Overview

  • Model: ChatOllama (e.g. llama3, granite3.3:2b)
  • Tools: Registered using langchain_mcp
  • Agent: Built with create_react_agent() from langgraph.prebuilt
  • Transport: Local streamable-http via FastAPI

📦 Setup

Install the required packages:

pip install langgraph langchain langchain-mcp langchain-ollama fastapi uvicorn

🛠️ Step 1: Define a Tool with MCP

Create a file called tools/weather_tool.py:

from langchain_mcp import tool, mcp
import os
import requests

@tool
def get_weather(location: str) -> str:
    """Fetch current weather for a given location."""
    api_key = os.getenv("WEATHER_API_KEY")
    url = f"http://api.weatherapi.com/v1/current.json?key={api_key}&q={location}"
    resp = requests.get(url)
    data = resp.json()
    return f"{location}: {data['current']['temp_c']}°C, {data['current']['condition']['text']}"

if __name__ == "__main__":
    mcp.run(
        transport={
            "type": "streamable-http",
            "bind": "127.0.0.1:8000",
            "path": "/mcp"
        }
    )

Run it:

python tools/weather_tool.py

🧠 Step 2: Build the Agent with LangGraph

Create a new file main.py:

from langchain_mcp_adapters.client import MultiServerMCPClient
from langchain_ollama import ChatOllama
from langgraph.prebuilt import create_react_agent
import asyncio
import os

async def main():
    client = MultiServerMCPClient({
        "weather": {
            "url": "http://localhost:8000/mcp",
            "transport": "streamable_http"
        }
    })

    tools = await client.get_tools()
    model = ChatOllama(model="granite3.3:2b")
    agent = create_react_agent(model, tools)

    response = await agent.ainvoke({
        "messages": [{"role": "user", "content": "What's the weather in London?"}]
    })

    for msg in response["messages"]:
        print(msg)

if __name__ == "__main__":
    asyncio.run(main())

✅ Result

The agent should:

  1. Parse the user query
  2. Select the get_weather tool via MCP
  3. Call the tool with location='London'
  4. Return the weather result in a natural reply

🧩 Notes

  • MCP allows tools to run independently in different processes or machines
  • LangGraph agents handle multi-step reasoning and looping if needed
  • You can define multiple tools (e.g., weather, math, search) and MCP will orchestrate them

🔗 Useful Links

Happy building!