This work presents a model predictive control scheme that provides offset-free setpoint tracking for systems described by a class of recurrent neural networks. The method is based on the reformulation of the model dynamics in velocity form, which embeds integral action in the closed loop and eliminates the need to compute the state and input equilibria associated with the setpoint. Terminal ingredients are designed via linear matrix inequalities, ensuring convergence and recursive feasibility under input and output constraints. The effectiveness of the approach is demonstrated through simulations on a nonlinear pH-neutralisation process.