I'm trying to build an MCP server that uses oauth and lets users perform CRUD operations on their data using Claude.

1. Is it correct that we should be creating a new MCP server (new McpServer) for every request? Every example does this, e.g. https://github.com/modelcontextprotocol/typescript-sdk/blob/main/examples/server/src/simpleStatelessStreamableHttp.ts. Is it safe to reuse the server instead?

2. When using oauth, how to access the auth context in a tool? E.g. get the user id. All of their examples have tools that don't depend on user data. The only solution I can think of is using AsyncLocalStorage, but I'd be surprised if there isn't an official way to do this.