The idea of brain machine interface (BMI) is to provide a source of interaction between a person and a machine via thought. Three major parts to an effective BMI are determined and handled in this paper: classifying a thought, doing a useful activity, and providing an efficient user interface (UI). This paper proposes an effective way of classifying thoughts and an approach for providing useful activities given a sequence of signals. We demonstrate the effectiveness of an extreme learning machine (ELM) for classifying a number of different thoughts when given a relatively small amount of training samples from a 5-channel EEG headset. We transform the electroencephalograph (EEG) data to a set of features for the input of the ELM model by estimating the logarithmic power (LP) of the discrete wavelet transform (DWT) coefficients, which corresponds to five different frequency bands. The ELM provides up to a 90% to 100% classification accuracy depending on training samples and number of hidden nodes, compared to 52% to 60% for a multi-layer perceptron (MLP). We also present a UI based on a state machine design that allows a person to accomplish certain activities via thought with a robotic arm.