Many interesting designs of soft robots with variablestiffness capabilities have been presented in the literature. However, little attention has been given on the control of their embedded physical intelligence. In this work, we present an algorithm that exploits the variablejoint stiffness capabilities and the redundancy of a soft manipulator to achieve Cartesian stiffness control at the end effector,thanks to model-based and optimal control techniques. The algorithm is validated both analytically and in the real world. In particular, we present a tendon-driven soft manipulator, equippedwith variable-stiffness segments and proprioceptive sensing. The robot is used as a platform to test the algorithm in real tasks, such as fitting a peg in the hole. Thanks to accurate modeling, the soft manipulator is able to obtain the desired stiffness at the end effector over the workspace.