Comment by JyB

Right. On a related note: "MCP is invisible to the LLM, or should be, under normal use". I am not convinced there are any downsides for the harness to explicitly feed MCP keywords into the LLM. It seem explicitly doing it has many advantages. For example, it allows the user to ask to interact with a resource through "the configured MCP" rather than directly through some local cli or HTTP endpoints; which can steer the LLM towards the proper actions. Otherwise, the model could in-theory be "clueless" with such a request.