Hackster Learning Series: AI Agentic Protocols

This GitHub repo complements The Learning Series articles on Hackster.io about AI Agentic Protocols. It covers:

• the Model Context Protocol (MCP) for dynamic tool and data access,
• the Agent2Agent Protocol (A2A) for direct agent collaboration,
• and the Agent Communication Protocol (ACP) for robust, interoperable communication.

The repo includes Python code to jump-start practical implementation of these protocols, offering immediate application opportunities.

📑 Table of contents:

• Environment Setup
• Part 1: Model Context Protocol (MCP)
• Part 2: Agent2Agent Protocol (A2A)
• Part 3: Agent Communication Protocol (ACP)
• Demo videos on YouTube

Environment Setup

1. Install the required Python packages, listed in the provided requirements.txt:

pip install -r requirements.txt

2. If using Azure OpenAI as your AI backend, set the following environment variables:

Variable	Description
AOAI_API_BASE	Base URL of the Azure OpenAI endpoint
AOAI_API_VERSION	API version of the Azure OpenAI endpoint
AOAI_DEPLOYMENT	Deployment name of the Azure OpenAI model

Part 1: Model Context Protocol (MCP)

This section demonstrates how an AI agent can dynamically discover and use external tools. The implementation uses an MCP Server (MCPServer_HomeAutomation.py) to expose home automation functionalities (tools) and an MCP Client (MCPClient_GradioUI.py) as a Gradio UI for user interaction.

1. Defining MCP Tools and Resources (MCP Server): functions decorated with @mcp.tool() or @mcp.resource() in the MCPServer_HomeAutomation.py file define callable actions (Tools) or retrievable data (Resources) for the AI.

   from mcp.server.fastmcp import FastMCP
   mcp = FastMCP("Home Automation")
   
   @mcp.tool()
   def control_light(action: str) -> str: # Generalized snippet
       # ...
       return "Light controlled."
   
   @mcp.resource("home://device_status")
   def get_device_status() -> str: # Generalized snippet
       # ...
       return "{}"
   
   if __name__ == "__main__":
       mcp.run() # Starts the MCP server

2. Establishing MCP Server Connection (MCP Client): the MCPClient_GradioUI.py starts the server as a subprocess and connects using MCPServerStdio to enable tool discovery.

   import subprocess
   from agents.mcp import MCPServerStdio
   
   server_process = subprocess.Popen([...]) # Start server
   
   mcp_server = MCPServerStdio(...)
   await mcp_server.__aenter__() # Initialize connection

3. Initialising AI Agent with MCP Servers (MCP Client): An Agent is initialised with the connected mcp_servers. The agent's instructions are dynamically updated based on MCP tool availability.

   from agents import Agent, OpenAIChatCompletionsModel
   
   agent = Agent(
       name="Home Assistant",
       instructions="Use tools...",
       model=OpenAIChatCompletionsModel(...),
       mcp_servers=[mcp_server], # Link MCP server
   )

4. Processing User Input and Running Agent (MCP Client): when a user inputs a query, Runner.run() is invoked. The AI model, aware of the MCP tools, decides whether to call a relevant tool or access a resource via the MCP layer to fulfill the request.

   from agents import Runner
   
   async def process_user_input(user_input, agent):
       result = await Runner.run(starting_agent=agent, input=user_input)
       return result.final_output

Note

Hackster article about MCP can be found here.

Part 2: Agent2Agent Protocol (A2A)

Caution

Work in progress. To be updated soon!

Part 3: Agent Communication Protocol (ACP)

Caution

Work in progress. To be updated soon!

Demo videos on YouTube

• AI Agentic protocols: Part 1 - MCP (Home Automation UI demo)
• Home Automation with MCP: Behind the Code