Environmental DNA (eDNA) analysis of water samples is on the brink of becoming a standard monitoring method for aquatic species. This method has improved detection rates over conventional survey methods and thus has demonstrated effectiveness for estimation of site occupancy and species distribution. The frontier of eDNA applications, however, is to infer species density. Building upon previous studies, we present and assess a modeling approach that aims at inferring animal density from eDNA. The modeling combines eDNA and animal count data from a subset of sites to estimate species density (and associated uncertainties) at other sites where only eDNA data are available. As a proof of concept, we first perform a cross-validation study using experimental data on carp in mesocosms. In these data, fish densities are known without error, which allows us to test the performance of the method with known data. We then evaluate the model using field data from a study on a stream salamander species to assess the potential of this method to work in natural settings, where density can never be known with absolute certainty. Two alternative distributions (Normal and Negative Binomial) to model variability in eDNA concentration data are assessed. Assessment based on the proof of concept data (carp) revealed that the Negative Binomial model provided much more accurate estimates than the model based on a Normal distribution, likely because eDNA data tend to be overdispersed. Greater imprecision was found when we applied the method to the field data, but the Negative Binomial model still provided useful density estimates. We call for further model development in this direction, as well as further research targeted at sampling design optimization. It will be important to assess these approaches on a broad range of study systems.