In energy-efficient buildings, the interactions and coupling effects between the building, its environment and its conditions of use play an important role on the energy balance. Recent development in information and communication technologies enable to have real time information about current and future environmental conditions, energy price or CO2 concentrations. Thus, it becomes possible to design building management systems that exploit these data in real time in order to optimize occupants comfort and energy performance. Model predictive control relies on a numerical model and real time measurements to compute an optimal strategy. This paper shows an example of application on a experimental building where both heating and ventilation are simultaneously controlled. Model predictive control is here computed so as to minimize a criterion based on operative temperature and energy consumption. The computation uses data collected on-site by temperature and power sensors, as well as weather data collected online through web-services. We describe here the deployment phase which includes a model calibration process in order to ensure optimality of the results. Both model calibration and optimal management are computed using a multizone thermal model. During optimization phases, the adjoint method is employed to derive efficiently descent algorithms. These choices make the whole software part of the system flexible and fast enough to be embedded on-board on a domestic building management system.