foyst/aws-knowledgebase-mcp-server

AWS Bedrock Knowledge Base MCP Server

With compliments to the following projects and their respective authors for providing the foundational work: https://github.com/daohoangson/aws-knowledge-base-mcp-server

https://github.com/modelcontextprotocol/servers-archived/tree/main/src/aws-kb-retrieval-server

Overview

This solution provides Retrieval Augmented Generation (RAG) capabilities using an AWS Bedrock Knowledge Base integrated with the Model Context Protocol (MCP). This can be integrated with various MCP-compatible AI agent clients, such as VS Code and Claude desktop, to provide private or specialised knowledge bases for additional context.

It's made up of two parts:

1. ./infra: The infrastructure code that sets up the Bedrock Knowledge Base, S3 bucket, IAM roles, and OpenSearch Serverless resources. This is deployed using Pulumi and SST.
2. ./mcp-server: The MCP server code that connects to the Bedrock Knowledge Base and serves as the MCP endpoint over http. This can be run locally through STDIO or deployed as a Lambda function.

Key mentions

• This uses https://github.com/CodeGenieApp/serverless-express, which seamlessly bridges Express.js applications to AWS Lambda allowing for easy deployment of the MCP server as a serverless function.
• The server is exposed using a lambda function URL for simplicity, but in a production scenario you would likely want to front this with an API Gateway or Application Load Balancer for better security and management.

How to use

1. Clone this repository.

2. Install dependencies for both the infra and mcp-server directories with npm install.

3. Configure your AWS credentials to allow Pulumi and SST to deploy resources.

4. Deploy the infrastructure using SST:

   npx sst deploy
   

5. Deploy the MCP server as a Lambda function or run it locally:

   To run it locally, you need to add it to your MCP configuration in the agent client of your choice. For example in VS Code, in the .vscode/mcp.json file add the following:

   {
     "servers": {
       "<name-your-mcp-client>": {
         "type": "stdio",
         "command": "node",
         "cwd": "${workspaceFolder}",
         "args": ["./mcp-server/dist/stdio.js"],
         "dev": {
           "watch": "mcp-server/**/*.ts",
           "debug": { "type": "node" }
         },
         "env": {
           "AWS_REGION": "us-east-1",
           "AWS_BEDROCK_KNOWLEDGE_BASE_ID": "<knowledge-base-id>",
           "AWS_ACCESS_KEY_ID": "<aws-access-key-id>",
           "AWS_SECRET_ACCESS_KEY": "<aws-secret-key>",
           "AWS_SESSION_TOKEN": "<aws-session-token>"
         }
       }
     },
     "inputs": []
   }
   

   You can get the AWS_BEDROCK_KNOWLEDGE_BASE_ID from the SST deployment output:

   

   If you want to connect to the remotely deployed lambda MCP server, you can add it as follows:

   {
     "servers": {
       "<name-your-mcp-client>": {
         "url": "<mcpServerUrl>",
         "type": "http"
       }
     },
     "inputs": []
   }
   

Debugging MCP Server locally with VS Code

Both the local and remote (lambda) MCP servers can be debugged using VS Code. For the local server with the debug property set to "node", you can hover over the server details in mcp.json select the "debug" option:

Then it's a case of adding breakpoints and invoking MCP requests from GitHub copilot (or other clients) as normal.

Known Limitations

• Despite using streaming, the response is not actually streamed back to the client, it is buffered and sent as a single response at the end.

Next Steps

• Provide authentication using OAuth 2.1 for the MCP server