Objective. Typically, a brain–computer interface (BCI) is calibrated using user- and session-specific data because of the individual idiosyncrasies and the non-stationary signal properties of the electroencephalogram (EEG). Therefore, it is normal for BCIs to undergo a time-consuming passive training stage that prevents users from directly operating them. In this study, we systematically reduce the training data set in a stepwise fashion, to ultimately arrive at a calibration-free method for a code-modulated visually evoked potential (cVEP)-based BCI to fully eliminate the tedious training stage. Approach. In an extensive offline analysis, we compare our sophisticated encoding model with a traditional event-related potential (ERP) technique. We calibrate the encoding model in a standard way, with data limited to a single class while generalizing to all others and without any data. In addition, we investigate the feasibility of the zero-training cVEP BCI in an online setting. Main results. By adopting the encoding model, the training data can be reduced substantially, while maintaining both the classification performance as well as the explained variance of the ERP method. Moreover, with data from only one class or even no data at all, it still shows excellent performance. In addition, the zero-training cVEP BCI achieved high communication rates in an online spelling task, proving its feasibility for practical use. Significance. To date, this is the fastest zero-training cVEP BCI in the field, allowing high communication speeds without calibration while using only a few non-invasive water-based EEG electrodes. This allows us to skip the training stage altogether and spend all the valuable time on direct operation. This minimizes the session time and opens up new exciting directions for practical plug-and-play BCI. Fundamentally, these results validate that the adopted neural encoding model compresses data into event responses without the loss of explanatory power compared to using full ERPs as a template.