Abstract
<jats:p>Large language model (LLM) agents are a promising route to automating rigorous process simulation, but the usual approach—having the model write the scripts that drive the simulator—is error-prone and insecure, especially for the compact open-weight models that industrial data confidentiality favors. This work instead connects LLMs to Aspen Plus® through the model context protocol (MCP), an open standard that lets an AI model operate external software through a predefined, validated set of operations rather than through code it writes itself. An MCP server exposes the simulator as a fixed set of strongly typed tools covering flowsheet synthesis, simulation, and analysis; requests that do not match a tool’s declared input format are rejected before they reach the simulator. A declarative builder constructs, converges, and analyzes a complete flowsheet from a single natural-language description, and the entire system runs on the user’s own hardware, so no process data leaves the site. Across 19 automatically scored tasks on three flowsheets, tool-calling raised the task success of a local 32B-parameter model from 16% to 89%, and of a 14B model from 0% to 84%; the gap persists under iterative error feedback, because the dominant failure of generated code—invented variable paths that do not exist in the simulator—cannot occur through typed tools. These results show that a standardized, schema-constrained interface makes compact local models viable for reliable, autonomous process simulation.</jats:p>