Protocols-MCP

Principle: Each tool should have a single, well-defined responsibility (Single Responsibility Principle) and provide clear, structured outputs that the agent can reason about.

Hosting MCP Server

The SDK wraps your agent functions as MCP-compatible servers, handling protocol details so you can focus on logic.
For tool hosting,
The SDK uses Stateless Streamable HTTP with the Mcp-Session-Id header for session isolation .
Servers must run on 0.0.0.0:8000 and expose /mcp.

AgentCore automatically adds Mcp-Session-Id to support stateless, session-aware communication. The InvokeAgentRuntime API passes through all payloads, enabling seamless MCP message proxying.

The Bedrock AgentCore Gateway is important because it streamlines the process of turning Lambda functions into secure, scalable MCP servers without managing infrastructure. 

Execution Flow
Agent Initialization: The agent starts and is configured to interact with the configured tools via the MCP Server hosted on AWS using the AgentCore Runtime.
Tool Discovery or Invocation: The agent sends a request to the MCP Server (using streamable-HTTP) to either list available tools or invoke a specific tool (e.g., add_numbers, multiply_numbers, or greet_users).
Tool Execution: The MCP Server, running inside the AgentCore Runtime, processes the request and routes it to the appropriate tool implementation.
Tool Response Returned: Once the selected tool completes its logic (e.g., performs a computation or generates a greeting), the MCP Server returns the response to the agent.
Agent Processes Result: The agent receives the result from the MCP Server and uses it to continue its reasoning or return a final response to the user or system.

Key MCP Server Components Explained

FastMCP: Creates an MCP server that can host your tools

@mcp.tool(): Decorator that turns your Python functions into MCP tools

stateless_http=True: Required for AgentCore Runtime compatibility

Tools: Three simple tools demonstrating different types of operations