Topic 2: The production app

What stays the same, and what gets added

notes_assistant/server.py already knows everything about MCP: tools, resources, the prompt, auth. None of that changes. Production concerns are about the process around it: how requests arrive, how many copies run, what happens under load, and how you see inside it. Keeping those in a separate file means Module 11's capstone can import the same build_server and choose how much of this to switch on.

A few terms first:

• ASGI app: a Python callable app(scope, receive, send) that a web server such as uvicorn drives. mcp.streamable_http_app() returns one; so does Starlette.
• ASGI middleware: an ASGI app that wraps another and can look at, change, or answer each HTTP request before the inner app sees it. It works on raw HTTP.
• MCP middleware: the SDK's own async (ctx, call_next) hook that wraps each decoded MCP message. It works on JSON-RPC methods. The two compose: ASGI middleware sees HTTP, MCP middleware sees MCP.
• Lifespan: startup and shutdown code for an ASGI app. The SDK's session manager must run inside it.
• Replica: one running copy of the server. A fleet is several replicas that serve the same URL through a load balancer.