Out of curiosity, I put this exact question to claude [0]. Here's a tl;dr of its answer with my refutations:

> MCPs expose tools with defined schemas, parameters, and return types.

CLIs do exactly the same thing

> Stateful Connections

A chat session with an LLM is exactly that!

> MCPs can return structured data (JSON, complex objects)

So can an agent with a CLI. E.g. they can just take output and > file.json - now they have a json file.

> MCPs can expose "resources" - like file contents, database schemas, or API documentation - that I can read directly

If you mention these in your prompt to an agent, then they know where to look and can access them too (and use keys etc as necessary)!

> MCPs can send progress updates, ask for clarification, or stream results.

So can a chat session with an agent.

> MCPs can implement fine-grained permission controls and rate limiting at the protocol level.

Rate limiting is easy for an agent. Fine-grained permission could be limited by the user of the agent (e.g. by giving a restricted key for the agent to use), so possible if desired.

tl;dr no added benefit whatsoever.

[0] https://claude.ai/share/4b339fbd-a6db-4fcb-86cd-0e8493aab663