The rapid advancement of large language models (LLMs) is reshaping artificial intelligence applications across scientific domains, including materials science. This paper presents the Multimodal Heat Treatment Research Agent (MHTR-Agent), a novel framework that harnesses the capabilities of LLMs to facilitate multimodal data integration and intelligent analysis in heat treatment research. Leveraging OpenAI's API and advanced natural language processing techniques, MHTR-Agent extracts critical features from unstructured sources such as textual reports, experimental logs, and image annotations, thereby uncovering underlying patterns in the "composition–microstructure–property" relationships. Using Cr13 martensitic stainless steel as a case study, the agent integrates metallographic images, process parameters, and elemental compositions to predict material hardness and wear resistance. The model achieves a mean absolute error (MAE) of 2.88 (HV), outperforming conventional machine learning baselines. Furthermore, a retrieval-augmented generation (RAG) assistant, MHT-Adv, enhances process parameter optimization while substantially reducing factual inconsistencies often observed in generic LLM outputs. This work not only improves the efficiency and accuracy of data-driven materials analysis but also exemplifies a scalable paradigm for AI-assisted scientific discovery and complex process design. Future directions include extending the framework's generalizability and incorporating continual learning strategies to adapt LLMs for dynamic industrial scenarios.