With the fast progress of quadruped robots, we also see the rise of advanced controllers able to take wholebody decisions without any model reduction. Recently, wholebody model predictive control have been demonstrated on several legged robots. Based on these results, this paper presents a novel walking controller. Contrary to previously demonstrated approaches, our controller does not require the pre-computation of guide trajectories for the foot, nor of specific foot location, but rather decides on the flight the best foot movement using an original cost formulation. The predictive controller is then applied at the actuator level using an impedance controller, without requiring the more costly low-level torque controller that previous methods used. The method is validated on the real robot Solo, using an open-source implementation based on the solver Crocoddyl. We evaluate in depth the quality of the produced walk, despite external disturbance, and provide longer experiments in the companion video.