Although large language models (LLMs) have transformed the way people can interact with information and receive personalized help in learning spaces, we have a limited understanding of how blind and visually impaired (BVI) learners use these LLMs-powered technologies. This paper explores the potential of LLMs to improve the BVI learners’ experience in EarSketch, a music coding platform designed to foster computational thinking through creative expression. Through co-design sessions with BVI learners and students, we identify key challenges such as intuitive search, navigation, and feedback, and propose LLM-based solutions to address these barriers. We synthesize key design implications with broader relevance for expressive computer science learning environments.