This paper proposes a general framework to control underwater vehicle-manipulator systems from external sensors. The design of the control law follows several constraints and criteria to optimize the overall behavior. The task is defined in the sensor space, that is suited for visual servoing from camera or sonar images. The control input distribution between the arm and the body is considered and separated between the approach and the intervention phases. The proposed framework is validated in a realistic simulation environment, imposing constant and varying disturbances at the velocity level.