Large language model orchestrated workflow integrating retrieval-augmented generation and model context protocol tools for urban-scale simulation-based resilience assessment

Cities face growing risks from compound heatwaves and electricity outages, yet applying resilience-oriented urban building energy modelling (UBEM) at urban-scale remains methodologically complex and labour-intensive. This study investigates how large language models (LLMs), combined with retrieval-augmented generation (RAG) and a custom Model Context Protocol (MCP) toolchain developed in this work, can support transparent, reproducible and scalable workflows for urban-scale building resilience assessment. A literature-grounded RAG layer is constructed from journal articles on UBEM and city resilience and queried through structured templates so that an LLM orchestrates the distillation of scenario settings, resilience metrics and aggregation strategies into machine-readable guidance. These outputs drive an MCP-enabled EnergyPlus toolchain that reconfigures and simulates an existing archetype-based UBEM for Shanghai, comprising 113 archetypes representing 311,612 buildings, under a three-day extreme heat event followed by a prolonged cooling outage. The workflow produces resilience indicators, including comfort-band occupancy and threshold exceedance statistics. At urban scale, the share of occupied hours within 20–26 °C falls from 42.5% to 20.9%, while hours above 35 °C increase from 0% to 26.7% under the outage scenario, with industrial and transport archetypes showing the largest additional warming. The results demonstrate that the proposed workflow can complete a LLM orchestrated scenario-to-indicator urban resilience assessment pipeline and reorganise expert effort from manual scripting towards governance of scenarios, thresholds and indicators, while clarifying methodological limitations and pathways for extending similar workflows to other cities, hazard configurations and UBEM applications.