Skills vs. Dynamic MCP Loadouts
3 days ago
- #Anthropic
- #Skills
- #MCP
- The author has transitioned from using MCPs (Microservice Control Protocols) to skills, including the Sentry MCP1.
- Discussions about dynamic tool loadouts and deferred loading of tool definitions have been ongoing, with Anthropic experimenting with wiring MCP calls via code.
- Deferred tool loading by Anthropic does not fully address the author's issues with MCPs, as tool definitions must still be static from the start of the conversation.
- Skills are summaries that inform the agent about available capabilities and how to use them, without loading full tool definitions into the context.
- The author tried using MCPs via CLI tools like mcporter but found it ineffective due to lack of tool awareness by the LLM and unstable MCP APIs.
- Manually maintained skills and agents writing their own tools have become the preferred method, offering more control and adaptability.
- The Sentry MCP, despite being well-designed, was abandoned due to high token costs and difficulties in integration.
- Future improvements may include dynamic tool loading with protocol changes to incorporate skill-like summaries and built-in manuals for tools.
- Protocol stability for MCPs is highlighted as a need, as frequent changes in tool descriptions disrupt materialized calls and external documentation.