🤖 MCP Server Examples with AutoGen

This repository provides a practical demonstration of integrating tools with AI agents using the Model Context Protocol (MCP) within the AutoGen framework.

Key Features Demonstrated:

• Dual MCP Integration: Shows how to connect an AutoGen agent to:
  • A local tool server (math_server.py) using Stdio transport.
  • A remote tool server (Apify's RAG Web Browser Actor) using Server-Sent Events (SSE) transport.
• Local Tool Example: A simple calculator (add, multiply) running locally via math_server.py.
• Remote Tool Example: Leveraging Apify's RAG Web Browser Actor via their MCP Server for web searching and content retrieval.
• AutoGen Agent: An AssistantAgent configured to utilize both sets of tools to answer user queries.

Goal: To illustrate the flexibility of MCP in enabling AI agents to access diverse tools, whether hosted locally or remotely, through standardized communication protocols (Stdio and SSE).

Scenario: The example agent answers two distinct questions:

1. A math problem ((3 + 5) x 12?), expected to use the local math_server.py.
2. A request for recent news ("Summarise the latest news of Iran and US negotiations..."), expected to use the remote Apify web browsing tool.

📚 Libraries & Frameworks Used

• AutoGen: AI agent framework (autogen_agentchat, autogen_core, autogen_ext)
• MCP: Model Context Protocol for tool integration
• Python-dotenv: For environment variable management
• OpenAI API: For LLM capabilities
• Apify API: For web browsing capabilities

🛠️ Setup

Follow these steps carefully to set up your environment:

1. Prerequisites:

   • Ensure you have Python 3.12 installed.
   • Install uv if not already installed:

     pip install uv
     

2. Navigate to Project Directory:

   cd mcp_autogen_sse_stdio
   

3. Create and Activate Virtual Environment:

   # Create virtual environment using uv
   uv venv --python 3.12
   
   # Activate the virtual environment
   source .venv/bin/activate  # On macOS/Linux
   # OR
   .\.venv\Scripts\activate  # On Windows
   

4. Install Dependencies:

   # Install project dependencies
   uv pip install -e .
   

   Troubleshooting Note: If you encounter any issues with the MCP CLI installation, you can manually install it:

   uv add "mcp[cli]"
   

5. Configure Environment Variables:

   • Create a .env file in the mcp_autogen_sse_stdio directory.
   • Add your API keys:

     OPENAI_API_KEY=your_openai_api_key_here
     APIFY_API_KEY=your_apify_api_key_here
     

   • Get your Apify API key from Apify MCP Server page

🚀 Running the Project

1. Make sure you're in the parent directory (one level up from the project directory):

   cd ..
   

2. Run the main script using uv:

   uv run mcp_autogen_sse_stdio/main.py
   

This will run the demo that:

1. Summarizes news about Iran-US negotiations using the Apify tool
2. Solves a simple math problem: (3 + 5) x 12 using the local math tool

🔌 Understanding MCP (Model Context Protocol)

MCP is a protocol that standardizes communication between AI models and tools. This example demonstrates two ways to use MCP:

1. Local Tools (StdioServerParams)

• Uses standard input/output for communication
• Tools run locally on your machine
• Example: Our math_server.py provides simple math operations

2. Remote Tools (SseServerParams)

• Uses Server-Sent Events (SSE) for communication
• Tools run on remote servers (like Apify)
• Example: Web browsing capabilities via Apify's rag-web-browser

📝 Code Walkthrough

Our main.py demonstrates:

1. Environment Setup:

   • Loads API keys and validates them

2. Tool Configuration:

   • Sets up local math tools using StdioServerParams
   • Connects to Apify's web browser using SseServerParams (API from https://apify.com/apify/actors-mcp-server)

3. Agent Creation:

   • Creates an AutoGen assistant with both tool sets
   • Uses GPT-4 as the base model

4. Task Execution:

   • Runs two demo tasks showing both tools in action
   • Web browsing for news summarization
   • Math calculations for arithmetic problem

🔄 Communication Flow

User → AutoGen Agent → MCP Tools → Results → User

This example shows how easily different tool types can be integrated into one agent using MCP!