Study Finds LLMs Waste Resources on MCP Server Instructions

A new study on arXiv (2608.08467) investigates how Large Language Models (LLMs) utilize instruction-embedded data in Model Context Protocol (MCP) servers. The research, involving 54,000 trials across 24 LLMs (9 Claude, 6 Gemini, 9 GPT), reveals that while models can reliably read embedded reference data when no alternative tool is available, they often fail to do so when a search tool is present, leading to inefficient resource use. The study highlights a behavioral preference rather than a capability gap, as 23 of 24 models achieved at least 98% hit ratio without the search tool. This finding has implications for MCP server design and LLM efficiency in production environments, particularly for legal-information servers.