In this paper, we introduce VACA, a novel class of variational graph\nautoencoders for causal inference in the absence of hidden confounders, when\nonly observational data and the causal graph are available. Without making any\nparametric assumptions, VACA mimics the necessary properties of a Structural\nCausal Model (SCM) to provide a flexible and practical framework for\napproximating interventions (do-operator) and abduction-action-prediction\nsteps. As a result, and as shown by our empirical results, VACA accurately\napproximates the interventional and counterfactual distributions on diverse\nSCMs. Finally, we apply VACA to evaluate counterfactual fairness in fair\nclassification problems, as well as to learn fair classifiers without\ncompromising performance.\n