This paper deals with the routing in cognitive mobile ad hoc networks. We propose to modify the widely adopted Ad-hoc On-demand Distance Vector (AODV) protocol [1] in order to assure its functionality in the considered scenario. The resulting protocol, referred to as the Cognitive Ad-hoc Ondemand Distance Vector (CAODV) protocol, has been designed according to three guidelines: i) to avoid regions of primary users activity during both route formation and packet discovery without requiring a dedicated common control channel; ii) to perform a joint path and channel selection at each forwarder to minimize the route cost; iii) to take advantage of the availability of multiple channels to improve the overall performances. The performances of CAODV have been evaluated by means of numerical simulations, and the experimental results confirm its effectiveness for cognitive mobile ad hoc networks.