We present a novel MPC method for humanoid gait generation that is guaranteed to produce stable CoM trajectories. This is obtained by using a dynamic extension of the LIP as motion model, with the ZMP velocity as a control variable, and embedding an explicit stability constraint in the formulation. Such constraint turns out to be linear in the control variables, leading to a standard QP problem with equality and inequality constraints. The intrinsically stable MPC framework is developed into a full-fledged gait generation scheme by including automatic footstep placement. Simulations show that the proposed method is very effective and performs robustly in the presence of changes in the prediction horizon.