We propose a model-agnostic stochastic predictive control (MASMPC) algorithm for dynamical systems. The proposed approach first discovers \textit{interpretable} governing differential equations from data using a novel algorithm and blends it with a model predictive control algorithm. One salient feature of the proposed approach resides in the fact that it requires no input measurement (external excitation); the unknown excitation is instead treated as white noise, and a stochastic differential equation corresponding to the underlying system is identified. With the novel stochastic differential equation discovery framework, the proposed approach is able to generalize; this eliminates the repeated retraining phase -- a major bottleneck with other machine learning-based model agnostic control algorithms. Overall, the proposed MASMPC (a) is robust against measurement noise, (b) works with sparse measurements, (c) can tackle set-point changes, (d) works with multiple control variables, and (e) can incorporate dead time. We have obtained state-of-the-art results on several benchmark examples. Finally, we use the proposed approach for vibration mitigation of a 76-storey building under seismic loading.