Designing MCP Tools: Improving AI Agent Performance

This article discusses the common challenges teams face when building MCP (Model Context Protocol) servers and how to design effective MCP tools. It introduces the Capability Triangle framework to understand the strengths and weaknesses of the three key parties involved: the LLM, the tool developer, and the domain expert.

Details

The article explains that the common perception of MCP being a failed technology is due to teams building MCP servers with too many tools, leading to agents struggling with tool selection. However, the author argues that the problem lies not with the MCP protocol itself, but with the way the tools are designed. Successful MCP implementations, such as those at GitHub and Block, have converged on a set of principles that improve agent performance, including using fewer tools, providing better descriptions, and focusing on outcome-oriented design. The article then introduces the Capability Triangle framework, which highlights the distinct strengths and weaknesses of the three key parties involved in MCP tool design: the LLM (the MCP client), the tool developer, and the domain expert. The LLM brings language understanding and reasoning capabilities but lacks domain knowledge and symbolic computation. The tool developer understands the technical capabilities of the tools but may not know the specific needs of the end users. The domain expert has deep knowledge of the problem domain but may not be familiar with the technical details of the tools. Effective MCP tool design requires balancing the needs and capabilities of these three parties, similar to how UX design principles like affordance, recognition over recall, and visibility of system status have been applied to improve human-computer interfaces over the years.