LLM-Powered Digital Twins for Interactive Urban Mobility Simulation: Integrating SUMO with AI Agents

Urban mobility simulation platforms such as SUMO are widely used in transportation research but remain technically demanding for non-experts and disconnected from modern AI capabilities. This work presents an LLM-powered, web-based transportation digital twin that integrates SUMO with the OpenAI Agents SDK and the Model Context Protocol (MCP) to enable natural language interaction, dynamic scenario editing, and AI-assisted decision support. Our framework allows users to conversationally specify simulation tasks, which are translated by an LLM agent into SUMO configuration and TraCI commands. We extend the simulation environment with new editing features including road network modifications, traffic signal retiming, parking supply and pricing, and event simulation (e.g., incidents, work zones, special events). Guardrails and handoffs ensure safe execution and transparent auditing of simulation changes. By incorporating external tools through MCP, the platform further enriches scenarios with geographic and contextual data from services like Google Maps. We demonstrate the system using Austin, TX as a case study, showcasing how LLM-augmented digital twins can support planners, policymakers, and researchers in testing interventions, evaluating resilience, and exploring sustainable mobility strategies. Our results highlight how coupling LLM with transportation simulation makes digital twins more intuitive, interactive, and deployable, advancing AI-assisted urban planning and decision-making.