Bedrock MCP Connector

What is Bedrock MCP Connector?

A TypeScript client for interacting with AWS Bedrock and MCP (Model Context Protocol) servers. It runs in Node.js environments and is intended for developers building AI-powered applications that combine Bedrock models with external tools.

How to use Bedrock MCP Connector?

Install via npm: npm install @juspay/bedrock-mcp-connector. Create a BedrockMCPClient instance with a model ID, region, optional MCP server URL, and system prompt. Connect to the MCP server (if URL provided), send prompts with sendPrompt, and disconnect when done. Use the CLI for interactive usage.

Key features of Bedrock MCP Connector

• Seamless integration with AWS Bedrock's Converse API
• Support for Claude and other Bedrock models
• Connect to MCP servers to discover and use available tools
• Register custom tool handlers with JSON Schema input validation
• Event-based architecture for streaming responses
• TypeScript support with full type definitions
• Command-line interface for interactive usage
• Modular design for easy integration into other projects

Use cases of Bedrock MCP Connector

• Build a chatbot that uses MCP tools to fetch real-time weather forecasts.
• Create an AI assistant that performs statistical calculations on user-provided data.
• Develop a research tool that searches Wikipedia through an MCP‑connected API.
• Embed tool‑augmented LLM logic into a larger Node.js application.

FAQ from Bedrock MCP Connector

What models does Bedrock MCP Connector support?

It supports Claude and other Bedrock models via the Converse API. The default example uses anthropic.claude-3-sonnet-20240229-v1:0.

How do I connect to an MCP server?

Provide the mcpServerUrl (e.g., http://localhost:5713/sse) when creating the client. Then call client.connect() to establish the connection.

Can I listen to streaming events?

Yes. Use client.getEmitter() to access an event emitter and listen for events like message, error, tool:start, tool:end, response:start, response:chunk, and response:end.

How do I register a custom tool?

Use the registerTool(name, handler, description, inputSchema) method. The handler is an async function that receives the tool name and parsed input, and must return an object with a content array.

Is there a command-line interface?

Yes. The README mentions a "command-line interface for interactive usage" as a feature, but no specific CLI commands are detailed.