It is a technically challenging problem to assess the instantaneous brain state using electroencephalography (EEG) in a real-time closed-loop setup because the prediction of future signals is required to define the current state, such as the instantaneous phase and amplitude. To accomplish this in real-time, a conventional Yule–Walker (YW)-based autoregressive (AR) model has been used. However, the brain state-dependent real-time implementation of a closed-loop system employing an adaptive method has not yet been explored. Our primary purpose was to investigate whether time-series forward prediction using an adaptive least mean square (LMS)-based AR model would be implementable in a real-time closed-loop system or not. EEG state-dependent triggers synchronized with the EEG peaks and troughs of alpha oscillations in both an open-eyes resting state and a visual task. For the resting and visual conditions, statistical results showed that the proposed method succeeded in giving triggers at a specific phase of EEG oscillations for all participants. These individual results showed that the LMS-based AR model was successfully implemented in a real-time closed-loop system targeting specific phases of alpha oscillations and can be used as an adaptive alternative to the conventional and machine-learning approaches with a low computational load.