MCP Server on Amazon Bedrock AgentCore Runtime (Minimal README)

This project demonstrates how to build, test, and deploy an MCP (Model Context Protocol) server onto Amazon Bedrock AgentCore Runtime with OAuth (Cognito) authentication. It also includes local and remote MCP clients for testing tool discovery and invocation.

---

Features

• MCP Server built with FastMCP

• Three example tools:

  • add_numbers
  • multiply_numbers
  • greet_user

• Local MCP client for testing

• Cognito-secured AgentCore Runtime deployment

• Remote client with automatic JWT refresh

• End-to-end demo: build → deploy → authenticate → invoke tools

---

Project Structure

mcp_server_project/
├── mcp_server.py                 # MCP server implementation
├── my_mcp_client.py             # Local testing client
├── my_mcp_client_remote.py      # Remote (Cognito-authenticated) client
├── invoke_mcp_tools.py          # Remote invocation demo (tool calls)
├── requirements.txt
└── notebooks/
    ├── hosting_mcp_server.ipynb                # Main tutorial notebook
    └── runtime_with_strands_and_bedrock.ipynb # Additional runtime examples

---

Local Development

Start the MCP server

python mcp_server.py

List available tools

python my_mcp_client.py

---

Deploy to AgentCore Runtime

Deployment is handled via the AgentCore Starter Toolkit, which:

• Generates a Dockerfile
• Builds and pushes container to ECR
• Creates an AgentCore Runtime
• Configures OAuth via Cognito

Run the deployment steps in the accompanying notebook:

• Hosting MCP Server on Amazon Bedrock AgentCore Runtime – OAuth Inbound Authentication

This notebook creates:

• Cognito User Pool + App Client
• JWT-based authorizer configuration
• AgentCore Runtime instance
• SSM + Secrets Manager entries for connection details

---

Remote Invocation

Once deployed, you can test the MCP server using:

python my_mcp_client_remote.py

Or invoke specific tools:

python invoke_mcp_tools.py

The clients automatically:

• Fetch Agent ARN from SSM
• Fetch Cognito tokens (and refresh when needed)
• Connect to the MCP server over streamable HTTP
• Perform initialize, list_tools, and call_tool operations

---

Requirements

• Python 3.10+
• AWS credentials configured
• Docker
• Amazon Bedrock AgentCore Starter Toolkit
• FastMCP & MCP libraries

Full dependencies listed in:

• requirements.txt

Notebooks reference:

• AWS Show & Tell session: AgentCore + Gateway deep dive

• AgentCore Runtime tutorial with Strands Agents

---

Cleanup

To remove deployed resources, use the cleanup section in the notebook:

• Deletes AgentCore runtime
• Removes ECR repo
• Removes SSM parameter + Secrets Manager entries
• always check manually via thee AWS console to ensure that the cleanup has been effective