CTNet combines a convolutional module (inspired by EEGNet) for extracting local and spatial features from EEG time series with a Transformer encoder module using multi-head attention to capture global dependencies in high-level features. A fully connected layer classifier then categorizes the EEG signals into motor imagery classes. The model achieved 82.52% accuracy on BCI IV-2a and 88.49% on BCI IV-2b datasets. It has been integrated into the braindecode toolbox.