Brain computer interface (BCI) for healthy people is a growing field. Minimizing the number of electroencephalography (EEG) channels is a key technological advantage for the application of BCI, which would make the system more mobile, easier to setup and long-time use in the real life. In this paper, to decrease the number of channels, multi-channel common spatial pattern (MCSP) algorithm is used to extract the features with two mental tasks (i.e., mental arithmetic and spatial imagery), and support vector machine (SVM) is used to classify the tasks performed. In detail, the separability value of each individual channel is computed based on between/within-group variance and a modified entropy criterion to evaluate its contribution to classification performance. The optimal channels are chosen based on the separability ranking. The performance of proposed methods is compared with recursive channel elimination and genetic algorithm. The results demonstrate that the EEG signals have different trends between the two mental tasks with highest brain activity in left central-parietal and parietal lobes, and the separability values allow reduction of number of electrodes from 15 to 4 and 10 while the classification accuracy reaches 80% and 90% respectively. Therefore, the optimal-channel algorithms can reduce the number of channels and improve the performance of the mental tasks based BCI.