You can see the observer as additional functonality to provide you additional (artificial) measurements. The input into the observer only can be signals you know, i.e., plant measurements and control inputs to the plant.

In your MPC scheme you can use then you measurements from the plant as well as the estimates of the observer. You can treat the observer outputs as additional measurements and are, in principal, free to choose. In case you need all of them in your MPC scheme take all of them, if you have a scheme which requires only a subset, only use a subset.

Adding on to what Daniel said, once you are using EKF, you can also use in your MPC scheme the linearized state matrices computed by the EKF, which reduces the computational burden.

Dear Daniel Ossmann , thank you for your answer. It was confusing to me because I thought inputs for observer are control inputs only and I may not choose plant measurements for inputs because I can't control them. Thank you.