Both, but for different reasons than most people argue.

CLI for defined workflows where the developer knows exactly what they want. MCP for discovery — when an AI agent needs to figure out what tools are available and how to compose them.

The real unlock with MCP isn't replacing CLI commands. It's that you can expose a tool once and every AI agent that supports the protocol can use it without custom integration. We built an MCP server that wraps our entire testing and deployment pipeline. Any Claude, Cursor, or Windsurf session can discover and invoke those tools without us writing a plugin for each IDE.

The CLI vs. MCP framing is wrong. CLI is for humans. MCP is for agents. You need both.

You are bot!