With deep learning (DL) development, EEG-based emotion recognition has attracted increasing attention. Diverse DL algorithms emerge and intelligently decode human emotion from EEG signals. However, the lack of a toolbox encapsulating these techniques hampers further the design, development, testing, implementation, and management of intelligent systems. To tackle this bottleneck, we propose a Python toolbox, TorchEEGEMO, which divides the workflow into five modules: datasets, transforms, model_selection, models, and trainers. Each module includes plug-and-play functions to construct and manage a stage in the workflow. Recognizing the frequent access to time windows of interest, we introduce a window-centric parallel input/output system, bolstering the efficiency of DL systems. We finally conduct extensive experiments to provide the benchmark results of supported modules. Our extensive experimental results demonstrate the versatility and applicability of TorchEEGEMO across various scenarios. •The first deep learning toolbox towards EEG-based emotion recognition.•A workflow that divides the recognition system into five plug-and-play modules.•Built-in functions cover datasets, transformations, models, algorithms, and more.•A novel window-centric EEG I/O is to enhance system effectiveness.•Experiments demonstrate benchmark performance across various scenarios.