Signal variability of electroencephalography (EEG)-based computer interfaces (BCIs), especially in motor imagery (MI) for rehabilitation, limits inter-subject generalization. Most MI-BCIs rely on intra-subject training, leading to long calibration sessions for each user. Even inter-subject transfer learning strategies, where large datasets are used to pretrain models, require substantial amounts of user-specific data to adapt and yield practical performance. Here, we present cross-subject backward optimal transport (XS-BOT), which extends backward optimal transport for domain adaptation to inter-subject transfer. Leveraging cued labels, XS-BOT aligns the features’ distribution of the target subject with the training features’ distribution, minimizing the amount of adaptation data and avoiding model retraining. XS-BOT was evaluated in two scenarios: cross-subject (multiple training subjects) and subject-to-subject (single training subject). For different base models, XS-BOT markedly outperformed the baselines using only 20 adaptation trials and three EEG channels. Cross-subject adaptation yielded accuracies similar to intra-subject setting, where a calibration session is needed. For subject-to-subject, results varied depending on the training subject, with cases that exceeded intra-subject results. By enabling accurate decoding with minimal calibration, XS-BOT moves MI-BCIs toward plug-and-play use in rehabilitation, supporting immediate feedback and longer therapy.