. @ProgramWithAi shared a talk about how MCPs are boring and it can be more powerful to just provide the LLM with a tool like eval().

It reminded me once again, how much we need to relinquish certain kinds of historic controls. These tools need different guards.