Compensatory movements are commonly observed post-stroke and can negatively affect long-term motor recovery. In this context, a system that monitors movement quality and provides feedback would be beneficial. In this study, we aimed to detect compensatory movements during seated reaching using a conventional tablet camera and an open-source markerless body pose tracking algorithm called MediaPipe [1]. We annotated compensatory movements of stroke patients per frame based on the comparison between the paretic and non-paretic arms. We trained a binary classification model using the XGBoost algorithm to detect compensatory movements, which showed an average accuracy of 0.92 (SD 0.07) in leave-one-trial-out cross-validation across four participants. Although we observed good model performance, we also encountered challenges such as missing landmarks and misalignment, when using MediaPipe Pose. This study highlights the feasibility of using near real-time compensatory movement detection with a simple camera system in stroke rehabilitation. More work is necessary to assess the generalizability of our approach across diverse groups of stroke survivors and fully implement near real-time compensatory movement detection on a mobile device.