Shared control methods distribute control between human operators and robots in demanding tasks, enabling collaboration that leverages their respective strengths and expertise. Sharing the task typically involves blending algorithms that combine human control inputs to (pre)planed assistance trajectories. Conventional blending techniques, such as Linear Blending, compute a combined output but neither guarantee feasibility of this shared motion, nor ensure compliance with safety or task-related constraints. This paper proposes to tackle feasibility and safety by formulating the blending strategy as the solution of a constrained optimal control problem, that enforces environment limits, task requirements, and physical capabilities. A Model Predictive Control approach is used to solve the optimization problem and anticipate constraints by predicting the robot motion over a receding time horizon. We evaluate this approach in simulated and real-world pick-and-place teleoperation experiments. The experimental study compares the Model Predictive Control approach to Linear Blending and full Teleoperation. The results show that the new framework offers significant improvements, as it provides a safer, more accurate, and repeatable response.