MCP or CLI? How to Choose Right Interface for Your AI Tools

The author argues that CLI-based tooling is generally a better default for AI agents because it’s more token-efficient, stateless, and broadly portable, while acknowledging MCP’s value in sandboxed clients, shared multi-agent sessions, and non-technical user distributions; examples include migrating Obsidian workflows to the CLI, Microsoft’s Playwright, Google’s gws, and the phyllotaxis OpenAPI helper.