Electrocorticography of the primary motor cortex (M1) is a promising tool for controlling a brain-computer interface (BCI). Electrocorticograms (ECoG) of the human M1 within the central sulcus (intrasulcal ECoG) have been rarely examined. In order to evaluate the usefulness of intrasulcal ECoG for BCI, we examined patients with subdural electrodes placed temporarily inside the central sulcus and over the sensorimotor cortex (gyral ECoG). Five patients were asked to perform or imagine two or three classes of simple upper limb movements. Univariate statistical analysis of the results revealed that the intrasulcal ECoG on M1 showed significant variability across movement classes. A support vector machine was used for classification of single-trial ECoG signals to infer movement class (neural decoding). The movement classes were predicted with 80-90% accuracy (chance level: 33% or 50%). To reveal the relative importance of anatomical areas for neural decoding, the decoding performance was compared between gyral and intrasulcal ECoGs. The intrasulcal ECoG on the motor bank showed higher performance than the equally-sized gyral ECoG or the intrasulcal ECoG on the sensory bank. Analysis using a short time window revealed that movement class could be decoded even before movement onset. These results suggest the usefulness of intrasulcal ECoG on M1 to infer upper limb movements and present a promising application for a practical BCI system.