Computer science teaching assistants (TAs) are vital for fostering student motivation and belonging, yet traditional training often lacks scalable, immersive practice to address challenges like low self-efficacy or excessive help-seeking behaviors, aiming to promote high self-efficacy and independence. Our developed Extended Reality (XR) and AI-based training tool addresses these challenges by simulating realistic TA-student interactions in a virtual office-hours environment. It employs a three-stage learning model—observe, interact, reflect—with five AI-driven student personas (e.g., high-need, low self-efficacy, solution-Focused) derived from instructor and TA consultations. TAs observe model interactions, engage with AI avatars displaying code snippets, and receive rubric-based feedback aligned with the MUSIC Model of Motivation, emphasizing empowerment, success, and caring, with interest and usefulness integrated indirectly. A formative usability test with nine participants (four faculty, two graduate TAs, and three undergraduate TAs) using Meta Quest 2 headsets praised the tool’s immersive environment and realistic personas for enhancing empathy and scaffolding skills. Issues like unclear stage transitions and verbose feedback prompted refinements, such as improved transition cues. As a work-in-progress, our tool shows promise for TA training, with plans for broader evaluation, refined feedback, and a collaborative mode for peer support. It offers a scalable, innovative approach to prepare TAs to enhance student motivation and learning outcomes in CS education, with potential for institutional adoption.