Talking to Cities: A Model Context Protocol Server for SUEWS

Large language models (LLMs) offer the potential to make complex scientific software accessible through natural language interfaces. However, LLMs hallucinate by design—generating plausible but incorrect physics, inventing parameter values, and confidently explaining non-existent model features. For scientific computing, this poses unacceptable risks to research integrity.

We present a solution: a Model Context Protocol (MCP) server for the Surface Urban Energy and Water balance Scheme (SUEWS). MCP, introduced by Anthropic in 2024 and now adopted by major AI providers, enables a fundamental architectural shift from AI-generated code to AI-orchestrated validated operations. Rather than prompting an LLM to write Python code and hoping it implements correct physics, we provide 15 typed tools with validated inputs and outputs. The AI can orchestrate these tools but cannot bypass validation or invent new operations.

The SUEWS-MCP server implements tools across five categories: configuration (create, update, validate, inspect), knowledge (list models, access schema, retrieve physics implementations), simulation (run SUEWS), utilities (calibrate OHM coefficients, document variables), and analysis (load and export results). Each tool enforces physical constraints—albedo must lie between 0 and 1, temperatures must exceed 0 K—rejecting invalid configurations before computation.

A key innovation addresses hallucination at the knowledge level. When explaining how SUEWS calculates storage heat flux, the AI retrieves and interprets actual Fortran source code rather than generating explanations from training data. If the implementation changes, the explanation changes. This direct coupling between AI responses and model code ensures trustworthy scientific communication.

We evaluated the system using 50 test questions across difficulty levels, comparing four configurations: baseline (no tools), reference (full repository access), and two MCP-enabled models. MCP improved answer accuracy by 18–20% over baseline, with largest gains on physics questions requiring equations and implementation details. The smaller model with MCP tools outperformed the larger model, demonstrating that tool access matters more than model size for domain-specific applications.

This work demonstrates that AI can make scientific software accessible without sacrificing rigour. Natural language interfaces become viable for urban climate modelling when AI orchestrates validated operations rather than generating unchecked code. The approach generalises: any computational tool with well-defined operations can expose an MCP interface, enabling trustworthy AI assistance across scientific domains.