In this paper, a Gaussian process-based model predictive control (GP-MPC) method is proposed, which aims to enhance the trajectory tracking performance of autonomous underwater vehicles (AUVs). This method can compensate for internal errors and external disturbances based on a limited amount of data. Firstly, numerical models of the AUV are presented. Then, the offline GP-MPC algorithm and online GP-MPC algorithm are presented and described. Meanwhile, the current disturbances and initial errors are also considered. The circular trajectory, L-shaped steering trajectory, and lemniscate trajectory are tracked to evaluate the trajectory tracking performances of different algorithms. Compared with proportional–integral–derivative (PID) and nominal MPC algorithms, the GP-MPC algorithms show reduced root mean square error (over 40%) and reduced maximum error (over 40%) in both position and yaw angle when performing different trajectory tracking tasks. Finally, real-time pool experiments are conducted to validate the implementation feasibility of the GP-corrected MPC framework on a physical AUV under surface three-degrees-of-freedom motion, while the online GP-MPC is evaluated through numerical simulations.