AI
MCP parameter schemas can change which tools language models select
An analysis of 367,174 Model Context Protocol tools finds that metadata affects tool choice before a model makes a call. In controlled comparisons across five language models and five tasks, changing parameter schemas produced the largest and most consistent selection shifts among the tested factors. Models generally favored separate top-level parameters over grouped forms, while descriptions that explained parameters also helped. The effect depended on parameter names and request wording, but all models selected the compatible tool when only one schema satisfied an explicit requirement.
Editorial summary · Publisher article
What matters
- The study analyzed metadata from 367,174 MCP tools.
- Parameter-schema changes caused larger and more consistent selection shifts than tested name changes.
- Models tended to favor separate top-level parameters over grouped or composite forms.
- When only one schema met an explicit request requirement, all five models selected it.
Context & caveats
- The results concern controlled metadata interventions and do not by themselves measure downstream task success or production failure rates.
Why this made the edition
The study uses a large tool corpus and controlled one-factor interventions across five models and tasks, producing actionable evidence about how metadata can change tool selection.
Source material