To handle the scarcity and heterogeneity of electroencephalography (EEG) data in Brain-Computer Interface (BCI) tasks, and to harness the vast public data, we propose Neuro-GPT, a foundation model consisting of an EEG encoder and a GPT model. The foundation model is pre-trained on a large-scale public EEG dataset, using a self-supervised task which learns how to reconstruct the masked chunk in EEG. We then fine-tune the foundation model on a Motor Imagery Classification task where only 9 subjects are available. Experiments demonstrated that applying foundation model can significantly improve classification performance compared to the model trained from scratch, which provides evidence for the advanced generalizability of foundation model and the ability to address the challenges of data scarcity and heterogeneity.