1. To investigate the structure of the primate burst generator for vertical saccades, we obtained intra-axonal records from vertical medium-lead burst neurons with upward on-directions (UMLBs) in alert, behaving squirrel monkeys, while monitoring their spontaneous eye movements. After physiological characterization, these UMLBs were injected with horseradish peroxidase. 2. UMLBs (n = 14) had no spontaneous activity and emitted bursts of action potentials that preceded rapid eye movements by approximately 6 ms. Parameters of the burst (duration and number of spikes) were highly correlated with parameters of the rapid eye movement (duration and amplitude of the upward displacement of the eyes). 3. The axons of six UMLBs projected to the oculomotor complex. Their somata (4 were recovered) were all in the rostral interstitial nucleus of the medial longitudinal fasciculus (riMLF). Their axons traveled caudally in the medial longitudinal fasciculus (MLF) and ramified in the interstitial nucleus of Cajal (NIC) before entering the oculomotor nucleus. Five axons terminated bilaterally in the subdivisions innervating the superior rectus and inferior oblique muscles and therefore were presumed to be excitatory. One axon terminated in the ipsilateral inferior rectus and superior oblique subdivisions of the oculomotor complex and was presumed to be inhibitory. 4. Additionally, our data demonstrate that the nucleus of the posterior commissure (nPC) may also contain UMLBs. The axon of one such neuron crossed the midline within the posterior commissure and provided terminal fields to the contralateral nPC, riMLF, NIC, and the mesencephalic reticular formation but not to the oculomotor complex. 5. In conclusion, our data demonstrate that the rostral mesencephalon of the monkey contains neurons that have both the activity and the connections that are necessary either to provide motoneurons innervating extraocular muscles of both eyes with the pulse of activity they display during upward saccades or to inhibit their antagonists. Furthermore, our data demonstrate that some UMLBs are better suited for closing the feedback path of the local feedback loop rather than for providing direct input to extraocular motoneurons.