The SCRIPT project aims at delivering machine-mediated hand and wrist exercises to people with stroke in their homes. In this context, adapting the exercise to the individual needs potentially enhances recovery. We designed a system composed of a passive-actuated wearable device, a personal computer and an arm support. The system enables users to exercise their hand and wrist movements by playing interactive games which were developed as part of the project. Movements and their required speed are tailored on the individual’s capabilities. During the exercise the system assesses whether the subject is in advance (leading) or in delay (lagging) with respect to a reference trajectory. This information provides input to an adaptive mechanism which changes the required movement speed in order to make the exercise neither too easy nor too challenging. In this paper, we show results of the adaptation process in a study involving seven persons with chronic stroke who completed a six weeks training in their homes. Based on the patterns observed in difficulty and lag–lead score, we defined five session types (challenging, challenging–then supporting, supporting, under-supporting and under-challenging). We show that the mechanism of adaptation has been effective in 195 of 248 (78.6%) sessions. Based on our results, we propose the lag–lead based assessment and adaptation as an auto-tuning tool for machine based exercise, with particular focus on rehabilitation robotics. Also, the classification of sessions among different types can be applied to other studies in order to better understanding the progression of therapy in order to maximize its outcome.