Towards Automated Map JSON Style from Spatial Vector Data Using MCP

Abstract. GIS data is inherently multi-dimensional, involving space, time, and attributes, and interpreting it usually requires considerable time and expertise. Many users of web-mapping platforms struggle at the visualization stage, where they must interpret the data, choose appropriate visualization methods, and define map elements such as size, colour, symbols, transparency, and overall composition. They must also understand map rendering through Map Style JSON, which typically demands significant technical knowledge and design experience. To reduce these barriers, we propose a system that combines the Model Context Protocol (MCP) with open Large Language Models (LLMs) served locally through Ollama to automatically generate Map Style JSON conformant with the MapLibre Style Specification directly from vector-based spatial data. The system is built around purpose-built MCP tools that inspect the data, build and validate data-driven styling expressions, and are designed to embed cartographic design principles such as semantic colour selection and the perceptually grounded use of visual variables. In a preliminary evaluation on consumer hardware, using a 77-polygon GeoJSON dataset and a vector tile service, the system recoloured a map's fill from a plain-language request and returned a style that MapLibre rendered without manual correction. The evaluation also shows that the local open model acts mainly as a natural-language front-end, while correctness and renderer compatibility depend on the MCP tool layer. Relying on open models running locally, the approach aligns with the open geospatial ecosystem and integrates with other open-source tools.