The digital transformation in online learning necessitates advanced systems that enhance learner engagement and provide personalized support. In this paper, we propose an AI tutor system integrated within a metaverse platform, to address these requirements. We design a four-layers architecture, including the user layer, functional layer, technology layer and data layer for the proposed system. The system utilizes a GPT API for generating context-aware responses, the Vosk API for speech recognition, the Vits API for high-quality voice synthesis, and the Unity Engine for scene rendering. In the experiment, twenty undergraduate and graduate students were recruited to use the AI tutor system for online learning. They were requested to take both pre- and post-test before and after the learning session. The result showed that a statistically significant improvement from pre-test to post-test scores (t = −10.68, p < 0.001), demonstrating potential for improving learning effects. Furthermore, we analyzed dialogue logs, and the result showed that high semantic alignment and logical consistency of AI tutor’s responses. Finally, we evaluated the user experience by a questionnaire. The result showed that 95% of subjects agreed that the AI tutor enhanced learning efficiency, and 85% found the interface easy to use.