Armin Ronacher has transitioned from using MCPs (Multi-Command Protocols) to skills, including moving away from the Sentry MCP1. He highlights discussions around dynamic tool loadouts that could delay loading tool definitions, but he remains skeptical about their effectiveness. Ronacher explains that while MCPs with deferred loading seem promising, they require significant engineering on the LLM (large language model) API side, which ultimately makes them less appealing compared to skills. Skills are concise summaries that provide the agent with essential information about available tools and how to use them. They don't load tool definitions into the context but help the agent understand how to utilize existing tools like bash. This contrasts with MCPs, where tool definitions must be static throughout the conversation. Ronacher notes that while it’s possible to invoke MCPs through the command line interface using tools like Peter Steinberger’s mcporter, this approach has limitations. The LLM struggles to recognize available tools, necessitating additional teaching to the model, which complicates matters. Ronacher finds that maintaining manual skill summaries for MCPs can become cumbersome, especially when MCP servers frequently change their tool definitions. He favors allowing agents to create their own tools as skills, providing greater control and adaptability. Although he acknowledges the potential for dynamic tool loading in the future, he emphasizes the need for protocol stability in MCPs. The ongoing changes in MCP tool descriptions hinder effective integration with external documentation and skills.