Brain-computer interface (BCIs) is non-muscular communication and control system which allows users to send messages or control commands to any kind of electronic devices via neuronal activity patterns. Magnetoencephalography (MEG) has a rich source about functional, physiological and pathological status of the brain and has widely used to investigate the neural activity of the brain. In this paper, a MEG-based BCI system was investigated using classification of brain activities during wrist movements in different directions. Performance of proposed method was evaluated on Dataset III of the BCI Competition IV. For each MEG channel, features were extracted in time-domain using some of the statistical properties of signal which are mean, standard deviation etc. At classification stage, Support Vector Machines (SVMs) and k-Nearest Neighbor (k-NN) algorithms are adopted. Classification was carried out among 'left', 'right', 'forward' and 'back' directions and experiments are conducted on per pair of these classes. Also, grid search and cross validation are adopted to achieve high classification performance. The best classification accuracy of 86.76% is achieved for Subject 1 using SVMs. It is believed that this study contribute to development of MEG-based BCIs.