OGSAgent: an MCP-based multi-agent collaboration framework for intelligent GIServices

Abstract

The vision of intelligent GIServices (Geographic Information Services) is to automatically discover, compose and execute distributed GIServices to address global challenges. Recent studies show that LLM (large language model)-based agents can autonomously use external services, with interoperability enabled by LLM capability server standards led by the Model Context Protocol (MCP). Achieving intelligent GIServices through LLM-based agents faces three issues: unified representation of geospatial capabilities, interoperability between geospatial and AI domains, and automatic coordination of heterogeneous GIServices. To address these issues, we propose OGSAgent, an MCP-based multi-agent collaboration framework for intelligent GIServices. First, it includes a GeoMCP server, which is a profile of the MCP server for the unified representation of geospatial capabilities. A common mapping method from the GIServices to the GeoMCP server is introduced and applied to OGC API-based GIServices. A multi-agent system is designed to automatically orchestrate geospatial capabilities, with agent roles and interactions following the ‘publish–find–bind’ GIService pattern and a purpose-based GIService taxonomy. Four use cases are demonstrated, including meteorological analysis, spatial analysis, remote sensing analysis and city digital twins. Ablation and comparative evaluations on a carefully designed multi-level benchmark dataset validate the effectiveness of OGSAgent for intelligent GIServices.