Most non-invasive brain-computer interfaces (BCIs) classify EEG signals. Here, we measured brain activity with magnetoencephalography (MEG) with an aim to characterize and classify single MEG trials during finger movements. We also examined whether averaging consecutive trials, or averaging signals from neighboring sensors, would improve classification accuracy. MEG was recorded in five subjects during lifting the left, right or both index fingers. Trials were classified using features, defined by an expert, from averaged spectra and time-frequency representations. Classification accuracy of left vs. right finger movements was 80-94%. In the three-category classification (left, right, both), accuracy was 57-67%. Averaging three consecutive trials improved classification significantly in three subjects. Instead, spatial averaging across neighboring sensors decreased accuracy. The use of averaged signals to find appropriate features for single-trial classification proved useful for the two-class classification. The classification accuracy was comparable to that in previous EEG studies. MEG provides another useful method to measure brain signals to be used in BCIs. Good performance was obtained when the classified signals were generated by two distinct sources in the left and right hemisphere. The present findings should be extended to multi-task cases involving additional brain areas.