Year: 2026
Type: article
Abstract: 3D city models in CityGML and CityJSON provide an important data foundation for urban analysis, yet they remain difficult to use in practice. Datasets are often large and heterogeneous, attributes may be missing, and many workflows require specialized domain knowledge. Large Language Models (LLMs) can lower this barrier by serving as natural language interfaces, provided that they interact with the underlying data structures reliably and consistently. This paper introduces CityDPC-MCP, a Model Context Protocol (MCP) server that wraps predefined functions for loading, querying, analyzing, and modifying 3D city models within a stateful dataset. The server is evaluated against two configurations available to users today: an LLM operating directly on the provided file and an LLM with code execution in a sandbox. Across five test cases, three models, and 100 repetitions each, the MCP setup achieved higher success rates while keeping the dataset out of the model context, which decouples token consumption from file size and format. An additional exploratory experiment on dataset generation illustrates possible agentic workflows and their current limitations.
Authors: Hamer, Tammo, Shamovich, Maxim
Institution: RWTH Aachen University
Language: English
Subfield: Building and Construction
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225425480