My company uses MCP in production to connect different AI products together that were developed by different teams. It makes the connections very easy, where a REST interface often has quirks and features that you have to understand.

That's fair. REST isn't much of a standard. MCP can be just as badly designed and documented as REST, but at least there's a standard for authenticating and listing the options. Reading through other top-level comments, this seems to be the theme. MCP usage is mostly driven by ease of connection.

Im not deep in there, but Id guess it offers some discovery mechanisms? (not sure if my understanding is correct)

Basically it's just bundling the API documentation into a standard format to plug into LLMs, if you're comfortable running prompt injection on yourself and downloading random text from random websites and telling your favorite LLM to use it to access your banking information.

It tells an agent what tools are available and the scenarios when they should be called. This does a ton of work when working with agents

It's a bit of a nothingburger IMO, but charitably you could think of it as a rest API + bundled agent instructions.