CLI or MCP? How should AI agents interact with external tools?

A comparison of two approaches for AI agents to interact with external tools: CLI tools and MCP (Model Context Protocol) servers. CLI tools offer simplicity, speed, and direct invocation with no setup overhead, making them ideal for stable, single-purpose tools. MCP servers provide a protocol layer enabling dynamic tool discovery, useful when managing many tools or building multi-agent systems. The key decision factor is whether your tool needs to be discovered dynamically or can be invoked directly. A practical heuristic: use CLI when the agent acts as yourself, use MCP when the agent acts on behalf of others.