There is quite a number of freely available EEG software for different paradigms (one such collection is MOABB - Mother of All BCI Benchmarks, and there’s a huge number of scientific articles).

The biggest bottleneck for a hobbyist is that when using EEG, most paradigms require somewhat expensive hardware to work and that most paradigms still don’t work well with scalp recordings outside a lab environment, even when using mid-cost devices.

There’s also the issue that classifiers usually have to be quite simple because datasets are small, because they are time consuming to record (and after you remove noisy epochs, you have even less data left). Cross-session and cross-subject learning rarely works, since EEG is dependent on so many factors like subjects’ brain anatomy, the type and precise location of electrodes, amount of gel (or lack thereof) and how dried out it is, mood and focus of the subjects, a huge number of environmental factors that influence subjects’ focus and many others.

The only paradigm I have seen to work a bit more reliably than others is Steady State Visual Potentials, because you have extra information that doesn’t need to be learned from EEG (the frequency of visual stimuli is roughly the same as the one in subjects’ occipital lobe).