It takes massive computation time to find the optimal hyperparameters of Gaussian Process. That can not be applied to the online training in real time applications or time-variant data source. The online algorithms proposed by other researchers are high computationally intensive. This manuscript presents natural gradient online algorithm for GP regression. GP may be used as a universal function approximator. Then an adaptive GP controller is designed in the state feedback control for a class nonlinear system. In order to demonstrate the availability of this adaptive GP controller, a simulation of the inverted pendulum system is given. The results of simulation demonstrate this GP online algorithm is very effective and the GP controller can achieve a satisfactory performance.