This paper addresses the problem of output feedback Model Predictive Control for stochastic linear systems, with hard and soft constraints on the control inputs as well as soft constraints on the state. We use the so-called purified outputs along with a suitable nonlinear control policy and show that the resulting optimization program is convex. We also show how the proposed method can be applied in a receding horizon fashion. Contrary to the state feedback case, the receding horizon implementation in the output feedback case requires the update of several optimization parameters and the recursive computation of the conditional probability densities of the state given the previous measurements. Algorithms for performing these tasks are developed.