If you use LLM CLI tools like Claude Code you can let model just call shell commands directly instead of MCP. Or does MCP have some advantage even in the scenario?