mcp-server-patterns

To build an MCP server with Node and TypeScript, you register tools, resources, and prompts using the @modelcontextprotocol/sdk. You must validate inputs with Zod and separate transport logic from business logic for reliable AI client communication.

The difference is that stdio is used for local MCP server connections while Streamable HTTP enables remote access. Choosing between them involves tradeoffs in network exposure, but keeping server logic transport-agnostic allows you to support both.

You need Zod schema validation in an MCP server to enforce structured input types for tools and resources. This schema-first approach prevents runtime errors by ensuring AI clients send properly formatted data before the server processes requests.

You register tools and resources using the evolving @modelcontextprotocol/sdk APIs by defining their schemas and handlers. Aligning with the current SDK versioning ensures your server properly exposes these capabilities for AI clients to invoke.

Best practices for production MCP servers include using schema-first validation, returning structured errors, ensuring idempotency, and applying rate and cost awareness. Maintaining SDK versioning and transport-agnostic logic are also required for production readiness.

Your MCP server fails to register tools after an SDK upgrade due to API changes in the @modelcontextprotocol/sdk. You must align your tool registration code with the current SDK versioning and debug transport or registration issues accordingly.