AssetOpsBench is a comprehensive benchmark and evaluation framework for AI agents operating in industrial asset management and predictive maintenance domains. The system provides a complete ecosystem: six domain-specific MCP (Model Context Protocol) servers exposing tools for IoT telemetry, time-series forecasting, work order lifecycle, failure mode analysis, vibration diagnostics, and utility functions; multiple agent runners (Claude, OpenAI, Deep, Stirrup, OpenCode, PlanExecute) that share a common trajectory model; and an offline evaluation pipeline with pluggable scorers (LLM-as-judge, deterministic static JSON, FMEA-specific, code-based). The architecture is MCP-native: all domain tools are exposed as FastMCP servers over stdio transport, and agents interact with them through a unified ToolUniverse client. Data is stored in CouchDB with config-driven loading, and the benchmark includes 141+ ground-truth scenarios across 9 asset classes, with sample data provided in-repo and the full corpus hosted on HuggingFace. The system is designed for reproducible agent evaluation, supporting both multi-agent orchestration and single-agent tool-use paradigms.