When we give LLM access to the terminal, LLM doesn't need anything else. LLM already knows all the commands, and there's no need to waste context on MCP. When we tell it to use a new command, it can get `--help` as needed.