Motivated by model predictive control of energy systems, we present a scalable nonlinear programming algorithm for dynamic optimization. The algorithm is based on a smooth exact penalty approach, coupled with a trust-region approach, and exhibits global convergence and local superlinear convergence while having excellent warm-starting properties so desirable in an online application. Moreover, the fact that it can achieve convergence entirely matrix-free recommends it for large scale approaches needing scalability. This builds on recent work of the authors where we proved using a generalized equations framework that such methods stabilize model predictive control formulation even when they have explicit inequality constraints. In particular, we present alternatives to enable fast active-set detection and matrix-free implementations.