Core concept: An MCP server exposes three kinds of capability: tools, resources, and prompts. The Python SDK (FastMCP) gives you one decorator per capability. The decorator names come straight from those capability names.

Why A and D are correct:

@mcp.tool registers a function the model can call. The MCP docs describe tools as "Functions that your LLM can actively call".
@mcp.resource registers read-only data the client can fetch. The docs describe resources as "Passive data sources that provide read-only access to information for context".
The SDK's own 15-line server uses exactly these two: @mcp.tool() on add, and @mcp.resource("greeting://{name}") on greeting.
Your type hints become the JSON Schema and the docstring becomes the description. You write no protocol code.

Why the others are wrong:

B: There is no @mcp.agent. A server exposes capabilities; it does not host agents. The agent lives in the client application.
C: There is no @mcp.action. "Action" is not an MCP term at all. The word for a callable function here is "tool".

Remember: MCP servers have three primitives, so three decorators: tool, resource, prompt. Anything else in the options is invented.

Question 174