Home-based rehabilitation with robot assistance may significantly reduce therapist workload and facilitate post-stroke patient training. However, the lack of real-time therapist supervision requires higher safety, ease of use, and training efficiency of the rehabilitation system. Therefore, active robot rehabilitation systems may not be particularly ideal for home use due to high costs, large footprints, safety concerns, and the potential for motor slacking. Self-powered robots provide a promising alternative by enabling the unimpaired side (UIS) to assist the impaired side (IS) without external power. However, constant and rigid coupling between limbs may lead to the over-compensation by the UIS, reduce engagement of the IS, and fail to provide sufficient information for task performance assessment. This paper demonstrates a home-based self-powered rehabilitation system with a variable stiffness mechanism (VSM), providing proper coupling stiffness between the UIS and IS according to the IS's motor ability. The application of the VSM also allows for the measurement of motion errors and interaction forces using only encoders, facilitating motor ability evaluation and training task guidance. First, we propose a cable-driven VSM with active and passive stiffness regulation, and develop a bilateral two-degree-of-freedom rehabilitation system with the VSM for evaluation. Then, the system's functionality and efficacy are validated through experiments with twelve healthy subjects and ten stroke survivors. The proposed self-powered system can potentially avoid UIS's over-compensation, enhances IS's engagement, and improves training efficiency, while ensuring safety and reliability within home-based rehabilitation.