In Model Context Protocol (MCP) systems, large language models (LLMs) choose tools based on metadata rather than on the underlying implementation. A tool’s name and description summarize its purpose, while the parameter schema specifies the fields the tool accepts and how they are organized. An analysis of 367,174 MCP tools shows substantial schema variation among functionally similar tools: many share the same name and similar descriptions but expose different parameter structures.
To test whether these schema differences can influence tool selection before any call is made, we performed controlled metadata interventions across five LLMs and five tasks. Each comparison altered exactly one factor—server name, tool name, description, or parameter schema—while keeping the others fixed. Tool positions were balanced, and conditional logistic regression was used to estimate position‑adjusted effects.
The findings are:
- Changes to the parameter schema produce the largest and most consistent shifts in selection. Models prefer separate top‑level parameters over grouped or composite forms.
- Descriptions that explain parameters tend to increase selection probability, whereas name effects are weaker and less consistent.
- Targeted comparisons reveal that schema effects depend on the specific parameter names and how the request represents the same information. When only one schema satisfies an explicit request requirement, all five models select the compatible tool despite its position in the list.
These results demonstrate that parameter schemas influence not only how tools are invoked but also whether they are chosen in the first place. For MCP developers and evaluators, schema design and tool ordering can affect which implementation is invoked even when functionality remains unchanged.
Review