MagicChat: An In-Tool AI Agent for Industrial Model-Based Engineering in MagicDraw

Abstract

Industrial model-based systems engineering (MBSE) relies on specialised modelling tools, yet engineers lack effective ways to leverage large language models (LLMs) inside these environments without externalising critical workflows. The key question is how to embed LLM agents into industrial modelling tools in a way that respects model semantics, preserves user control, and handles engineering artefacts. Prior work does not combine in-tool LLM integration, tool-mediated model operations with transactional safety, and structured document ingestion into a cohesive approach adaptable across MBSE tools and domain-specific modelling languages (DSMLs). We present an architectural approach that operationalises LLM agents as a model-native interaction mode, implemented as MagicChat, a native MagicDraw plugin providing context-aware model querying, transactional model manipulation with undo, large-document ingestion, and document generation. We validate it on a deployed medical-device model through a cumulative four-act scenario executed with three off-the-shelf LLMs, combining a structural analysis of the resulting artefacts with an independent review by 15 system architects. The architecture proved LLM-agnostic, every element the agent created conformed to the deployed DSML profile, and tool-call volume did not predict the quality of the result.