AMQ-Bench: Agent-Oriented MCP Quality Benchmark

AMQ-Bench is an agent-oriented benchmark for evaluating the quality of Model Context Protocol (MCP) services, centered on the Repository-to-MCP-Service task. It operationalizes Availability, Usability, and Utility over a corpus of 269 repository-to-service instances drawn from 126 GitHub repositories across four sources (Self-built, ToolArena, ScienceAgentBench, ResearchCodeBench). The release includes the full corpus (amq_bench.jsonl, 269 instances), a mini30 development subset (30 instances), a Croissant 1.0 metadata file, and a dataset card describing the evaluation framework and responsible-AI considerations.

This is an anonymous release accompanying a NeurIPS 2026 Evaluations and Datasets Track submission; author and institutional information will be added after the review period.