In Model Context Protocol (MCP) systems, an LLM never sees a tool's implementation; it chooses which tool to call based on metadata. That metadata typically includes a name, a description, and a parameter schema. A new arXiv preprint (2610.14334) asks which of these components actually drive correct selection.
The authors conduct controlled factor-level interventions, varying individual metadata fields while holding others fixed. This lets them isolate the contribution of each component rather than just correlating metadata quality with overall performance.
The abstract is truncated, so the specific results are not available in the source. The paper's framing, however, highlights a practical problem: if names and descriptions are what matter most, then MCP tool authors should invest there; if the schema is decisive, then schema design deserves more attention.