Author(s): Meagan Tobias; Jacob Marchionda; Branko Kerkez
Linked Author(s):
Keywords: Artificial intelligence; Decision support system; Hydrological modeling; Model Context Protocol; Urban drainage systems; Workflow automation
Abstract: Recent advances in large language models (LLMs) offer new opportunities to make water management more intuitive through natural language interfaces. However, the generalized training of these models limits their effectiveness at domain-specific tasks such as stormwater modeling, particularly when nuanced answers require executing and modifying water models. The Model Context Protocol (MCP) addresses this challenge by providing a standardized interface for LLMs to interact with external tools while enforcing guardrails for quality control. Building on this protocol, we introduce SWMM-MCP, a cross-platform framework, model agnostic framework that enables conversational interaction with the United States Environmental Protection Agency’s Stormwater Management Model (EPA SWMM) SWMM-MCP leverages the SWMM-API and pySWMM libraries through an MCP server built with Fast-MCP, allowing users to query, edit, run, and visualize models via natural language. We demonstrate SWMM-MCP using a benchmark model, evaluating how three different LLM configurations perform when tasked with comparing pipe performance under different storm scenarios. All three models successfully completed the task but required varying numbers of prompts (2-7) and exhibited different token usage patterns and costs ($0.06-$0.15). Results highlight that even within a structured MCP framework, LLM selection significantly affects task efficiency, and domain-specific prompting remains important for optimal performance.
Year: 2026