Evals

How we test MCP servers whose caller is an LLM — and what we deliberately didn't build.