Claude just reads the SKILL.md frontmatter into initial context and when the instructions match it reads the rest of the SKILL.md. Every LLM can do that.

If it's such a thing that every model already can support it and you don't need to do anything with it, why is it even a feature?

To be fair other providers have to implement reading a skill directory and only the frontmatter of the SKILL.md and read more of it when needed. But it's a LOT simpler than implementing MCP.