A prominent AI developer argues that Model Context Protocol (MCP) remains superior to the emerging Skills standard for connecting LLMs to services. While Skills work well for pure knowledge transfer, MCP's architectural approach of providing direct API access proves more practical for real-world integration scenarios.

The critique centers on Skills' reliance on local CLIs, which creates deployment headaches and authentication nightmares. Many environments like ChatGPT or Perplexity can't run arbitrary commands, making Skills-dependent tools unusable. MCP sidesteps these issues with remote servers that handle authentication through OAuth and provide seamless updates without requiring local installations.

For services developers want to integrate, the author recommends building remote MCP servers rather than Skills that instruct LLMs to install CLIs. The vision calls for native MCP endpoints from major platforms like Google Calendar, Chrome, and Notion, while reserving Skills for teaching workflows and organizational knowledge. This architectural split would create a cleaner ecosystem where connectors handle service integration and manuals handle contextual guidance.