Hi! Prodiy expects you do define the label set you want to use for annotation upfront, before you start the task.
If your goal is to train a model, every label you add will have an effect on the entire dataset, including previously created annotations. It impacts how the model learns to predict all other categories, and of course it could mean that previously assigned labels are now incorrect and belong to the new category. (That's also why designing the label scheme is such an important part of developing a new model. It's also intertwined with the machine learning and development side of things and not something that can be left entirely to the annotator – which is also why having an option to add labels in the UI would be counterproductive.)
If you're annotating and you notice that you missed a class, you should probably stop and restart the server – and possibly even revisit previously created annotations to make sure there are no unintended side-effects, inconsistencies and conflicts caused by the newly added label. These problems are often quite sublte and difficult to debug later on once you've collected a large dataset.