Is this what MCP 2.0 is all about?

The core of MCP 2.0 is:
LLM does not see tool definitions
LLM only generates code
The new sandbox is what MCP 1.0 is called

In MCP 2.0, the pipeline is: LLM → generate code → sandbox → gateway → MCP 1.0.
Does this mean the LLM must be my own LLM (either a paid API like Claude/OpenAI or a local model), and that the IDE-built-in LLM from Cursor/Trae cannot be used for MCP 2.0?