In this letter, we develop an optimal control framework that takes the full-body dynamics of a humanoid robot into account. Employing full-body dynamics has been explored in, especially, an online optimal control approach known as model predictive control (MPC). However, whole-body motions cannot be updated in a short period of time due to MPC's large computational burden. Thus, MPC has generally been evaluated with a physical humanoid robot in a limited range of tasks where high-speed motion executions are unnecessary. To cope with this problem, our multi-timescale control framework drives whole-body motions with a computationally efficient hierarchical MPC. Meanwhile, a biologically inspired controller maintains the robot's posture for a very short control period. We evaluated our framework in skating tasks with simulated and real lower-body humanoids that have rollers on the feet. Our simulated robot generated various agile motions such as jumping over a bump and flipping down from a cliff in real time. Our real lower-body humanoid also successfully generated a movement down a slope.