Upper extremity impairments are common among stroke survivors. Robotic devices enable a high-dose of repetitive training for patients, but most systems are confined to the laboratory settings due to their complexity and power requirements. Previously we developed a passive elbow device that can counteract the angle-dependent tone of flexor muscles with hypertonia, but its efficacy was found limited as the increase in passive assistance during elbow extension was found not sufficient to provide assistance to those with more severe impairments. Therefore, in this study, we developed a 'self-adaptable' passive device that adjusts its assistance level based on the movements of patients. In addition to the morphological design to adjust moment arms of the elastic components, we incorporated a self-adaptation mechanism, in which the lengths of the elastic bands were adjusted by a pair of miniature linear motors based on the joint position feedback signals. The capacity of the device was then tested in a pilot testing with two healthy subjects, for whom angle-dependent flexion torque was implemented to simulate flexor hypertonia. The additional adjustment of passive component lengths was found to further increase the elbow extension assistance as the elbow joint extended. The proposed self-adapting mechanism, which does not require any complex control input from the experimenters, can be incorporated with the existing passive device to improve its functional efficacy in home-based training.