PulseAugur
EN
LIVE 18:06:29

New tool reveals non-deterministic AI routing in reference server

A new tool called routeproof has been developed to test the reliability of AI model routing to specific tools. When tested against the @modelcontextprotocol/server-filesystem reference implementation, routeproof revealed that the AI model's tool selection is not deterministic, with results varying even with identical inputs. The tool identified a specific issue where the `list_allowed_directories` function was incorrectly called for file reading and directory tree requests, despite its description not encompassing those functionalities. The developer's initial prediction that a deprecated `read_file` tool would be the source of the error was disproven, highlighting the subtle nature of routing failures. AI

IMPACT This tool could improve the reliability of AI agents by identifying and fixing routing errors in tool selection.

RANK_REASON The item describes a new tool for testing AI routing, not a core AI model release or research.

Read on dev.to — MCP tag →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

New tool reveals non-deterministic AI routing in reference server

COVERAGE [1]

  1. dev.to — MCP tag TIER_1 English(EN) · Hex ·

    I pointed my routing tester at a server I didn't build. One tool was quietly stealing two others' jobs.

    <p>A few weeks ago I built a small tool called <a href="https://github.com/tamasPetki/routeproof" rel="noopener noreferrer">routeproof</a>. It tests one narrow, annoying thing about MCP servers: when an AI host decides which of your tools to call, the only thing its model sees is…