It depends on the measurement method. In general, computer should have ability for distinguish fast and reproducible mental states. For EEG it's mostly imagination of movements and strong emotions, for EP - readiness states, signals recognizing.
Besides on the comments that are going to be posted, I would consider that feedback and response times are important metrics in the implementation of an effective BCI. You do not want to have something that reacts almost real-time with the use of a system that can only provide slow action, remember that the brain activity in EEG is measuring the activation of a certain group of neurons, located at a specific location based on specific stimuli and it takes an important amount of time to achieve that with untrained subjects.
In such sense, the feedback method has to be selected carefully based on the previous literature reviews. The response has to be corrected from any errors based also in the measuring methods. EEG is most common for low-cost research but I highly doubt you will end up doing MEG, Spikes or ECoG without the participation of a big laboratory.