Embodied cognition theories have generated considerable interest in the last two decades because they claim to have solved the problem of concept representation in the human brain. The idea is that concepts are constituted by information that is represented in the sensorimotor system. Finger counting offers a unique paradigm to test this idea because it associates abstract numerical symbols with a physical counterpart which is useful for learning to count, which retains conceptual properties such as ordinality or cardinality, and which can be evoked through subtle – possibly covert – movements. We brought together data from developmental, neuroimaging and electrophysiological studies and confronted them to the idea that finger counting may lead to embodied numerical representations. Our review shows that, while evidence for number and finger interactions has accumulated, inferences are limited due to the paucity of studies contrasting competing hypotheses. To foster theoretical progress, we propose an original framework articulated around the questions of the necessity and sustainability of embodied representations. We draw four hypotheses that differ according to whether finger counting is viewed as ancillary or necessary to the development of efficient numerical skills and whether it serves numerical acquisition only or sustains numerical cognition over the lifespan through the re-enactment of sensorimotor experience. To discriminate between these hypotheses, we argue that future research should turn to causal evidence. Among the new directions, we explore the opportunities offered by the study of congenital sensorimotor disorders or the possibility to disrupt the body schema by inducing sensorimotor conflict.