Patients with tremor can benefit from wearable robots managing their tremor during daily living. To achieve this, the interfaces controlling such robotic systems must be able to estimate the user's intention to move and to distinguish it from the undesired tremor. In this context, analysis of electroencephalographic activity is of special interest, since it provides information on the planning and execution of voluntary movements. This paper proposes an adaptive and asynchronous EEG-based system for online detection of the intention to move in patients with tremor. An experimental protocol with separated self-paced wrist extensions was used to test the ability of the system to detect the intervals preceding voluntary movements. Six healthy subjects and four essential tremor patients took part in the experiments. The system predicted 60 ± 10% of the movements with the control subjects and 42 ± 27% of the movements with the patients. The ratio of false detections was low in both cases (1.5 ± 0.1 and 1.4 ± 0.5 false activations per minute with the controls and patients, respectively). The prediction period with which the movements were detected was higher than in previous similar studies (1.06 ± 1.02 s for the controls and 1.01 ± 0.99 s with the patients). Additionally, an adaptive and fixed design were compared, and it was the adaptive design that had a higher number of movement detections. The system is expected to lead to further development of more natural interfaces between the assistive devices and the patients wearing them.