Langchain MCP Connect

Introduction

This project introduces tools to easily integrate Anthropic Model Context Protocol(MCP) with langchain. It provides a simple way to connect to MCP servers and access tools that can be made available to LangChain. Most importantly, langchain-mcp-connect allows developers to easily integrate their LLMs with a rich ecosystem of pre-built MCP servers.

MCP integrations with langchain expands the capabilities of LLM by providing access to community servers and additional resources. This means that we do not need to create custom tools for each LLM, but rather use the same tools across different LLMs.

Installation

pip install langchain-mcp-connect

What is the Model Context Protocol (MCP)?

The Model Context Protocol (MCP) is an open-source standard released by Anthropic. The Model Context Protocol highlights the importance of tooling standardisation through open protocols. Specifically, it standardises how applications interact and provide context to LLMs. Just like how HTTP standardises how we communicate across the internet, MCP provides a standard protocol for LLM to interact with external tools. You can find out more about the MCP at https://github.com/modelcontextprotocol and https://modelcontextprotocol.io/introduction.

Example usage

The langchain_mcp_connect contain key methods to connect MCP server tools to LangChain. Before starting, please ensure you meet the pre-requisites. For a detailed example on how langchain_mcp_connect can be used, see this demo project.

Defining a tool

Define your tool within claude_mcp_config.json file in the root directory. For a list of available tools and how to configure tools see here.

langchain_mcp_connect supports both stdio and HTTP with Server-Sent Events (SSE) protocols.

{
  "mcpServers": {
    "git": {
      "command": "uvx",
      "args": ["mcp-server-git", "--repository", "./"]
    },
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "./"
      ]
    },
    "github": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-github"
      ],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ENV_GITHUB_PERSONAL_ACCESS_TOKEN"
      }
    },
    "sseService": {
      "url": "http://localhost:8000/sse"
    }
  }
}

Environment Variables

Managing secrets is a key aspect of any project. The langchain_mcp_connect tool is able to inject secrets from the current environment. To do so, prefix the name of your environment variable with ENV_ in claude_mcp_config.json to inject environment variables into the current context. In the example above, ensure you have defined GITHUB_PERSONAL_ACCESS_TOKEN in your current environment with:

export GITHUB_PERSONAL_ACCESS_TOKEN="<YOUR_TOKEN_HERE>"
export OPENAI_API_KEY="<YOUR_KEY_HERE>"

Running the example

You can find an example usage in the src/example/agent.py file. You will need to install uv.

uv run src/example/agent.py

Example code

import asyncio
import logging
import os

from dotenv import load_dotenv
from langchain_core.messages import HumanMessage
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent

from langchain_mcp_connect import LangChainMcp

load_dotenv()

logging.basicConfig(level=logging.INFO)
log = logging.getLogger("LangChainMcp")

if "GITHUB_PERSONAL_ACCESS_TOKEN" not in os.environ:
    raise ValueError(
        "Please set the GITHUB_PERSONAL_ACCESS_TOKEN environment variable."
    )
if "OPENAI_API_KEY" not in os.environ:
    raise ValueError("Please set the OPENAI_API_KEY environment variable.")


if __name__ == "__main__":

    QUERY = "What tools do you have access to?"

    # Define the llm
    llm = ChatOpenAI(
        model="gpt-4o",
        model_kwargs={
            "max_tokens": 4096,
            "temperature": 0.0,
        },
    )

    # Fetch tools
    mcp = LangChainMcp()
    tools = mcp.list_mcp_tools()

    # Bind tools to the agent
    agent_executor = create_react_agent(llm, tools)
    human_message = dict(messages=[HumanMessage(content=QUERY)])
    
    # Run the agent asynchronously
    response = asyncio.run(
        agent_executor.ainvoke(input=human_message)
    )

    log.info(response)