Visual inputs to the brain are mapped in a retinocentric reference frame, but the motor system plans movements in a body-centered frame. This basic observation implies that the brain must transform target coordinates from one reference frame to another. Physiological studies revealed that the posterior parietal cortex may contribute a large part of such a transformation, but the question remains as to whether the premotor areas receive visual information, from the parietal cortex, readily coded in body-centered coordinates. To answer this question, we studied dorsal premotor cortex (PMd) neurons in two monkeys while they performed a conditional visuomotor task and maintained fixation at different gaze angles. Visual stimuli were presented on a video monitor, and the monkeys made limb movements on a panel of three touch pads located at the bottom of the monitor. A trial begins when the monkey puts its hand on the central pad. Then, later in the trial, a colored cue instructed a limb movement to the left touch pad if red or to the right one if green. The cues lasted for a variable delay, the instructed delay period, and their offset served as the go signal. The fixation spot was presented at the center of the screen or at one of four peripheral locations. Because the monkey's head was restrained, peripheral fixations caused a deviation of the eyes within the orbit, but for each fixation angle, the instructional cue was presented at nine locations with constant retinocentric coordinates. After the presentation of the instructional cue, 133 PMd cells displayed a phasic discharge (signal-related activity), 157 were tonically active during the instructed delay period (set-related or preparatory activity), and 104 were active after the go signal in relation to movement (movement-related activity). A large proportion of cells showed variations of the discharge rate in relation to limb movement direction, but only modest proportions were sensitive to the cue's location (signal, 43%; set, 34%; movement, 29%). More importantly, the activity of most neurons (signal, 74%; set, 79%; movement, 79%) varied significantly (analysis of variance, P < 0.05) with orbital eye position. A regression analysis showed that the neuronal activity varied linearly with eye position along the horizontal and vertical axes and can be approximated by a two-dimensional regression plane. These data provide evidence that eye position signals modulate the neuronal activity beyond sensory areas, including those involved in visually guided reaching limb movements. Further, they show that neuronal activity related to movement preparation and execution combines at least two directional parameters: arm movement direction and gaze direction in space. It is suggested that a substantial population of PMd cells codes limb movement direction in a head-centered reference frame.