In MCP tool choice, the parameter schema decides: rename one field and selection drops from nine in ten to one in ten

When MCP tools do the same job, the parameter schema, not the server name or tool name, decides which one a model picks. Earlier work on tool-selection bias focused on names and descriptions, and parameter schemas were rarely varied on their own. Across three-tool comparisons on five tasks and five models, 23 of 28 numerical estimates had 95% confidence intervals excluding zero and 22 of those were positive; the schema coefficient was the largest in 19 of 30 task-model conditions, with a median magnitude of 3.82 against 1.48 for descriptions.

Models prefer parameters laid out as separate top-level fields over the same inputs packed into one object or dictionary. In ecosystem metadata, tools with the same name and similar descriptions differ widely in how inputs are written: 169 records named send_email used 78 distinct parameter sets. In Create Issue, all five models selected the tool with separate top-level parameters in 93.3–100% of trials; in Create Pull Request, all five favored separate owner and repo fields when the request gave them separately, and four still did when the request gave a combined owner/repo path.

Parameter names change whether a model judges a tool to fit the request, not just how arguments are filled in after selection. The parameter schema had been treated mainly as a specification of how a tool is invoked rather than as a factor in whether it is chosen. Renaming to as recipient_id lowered the renamed tool's selection rate to 1.0–10.0% across five models, while renaming it to receiver, a word absent from the requests, kept it at 35.2–51.0%; when a request asked for a CC recipient that only one schema could accept, all five models selected that tool even when it was listed second.

The findings cover the five tested models, five tasks, the constructed metadata variants, and candidate sets of two to five tools, and are limited to single-turn, LLM-generated English requests without tool execution. The intended readers are MCP server developers, host designers, and tool-selection evaluators; the experiments reproduce only the presentation and selection steps and execute no tool.

The authors state that they measure selection behavior rather than decision processes, so their reading of naming effects, such as recipient_id signaling an identifier rather than an email address, remains behavioral. The findings are limited to the tested models, tasks, metadata variants, and candidate-set sizes, and because the evaluated models are accessed through APIs, exact results may differ in later runs.