Up to now, the execution of ATL transformations has always followed a two-step algorithm: 1) matching all rules, 2) applying all matched rules. This algorithm does not support incremental execution. For instance, if a source model is updated, the whole transformation must be executed again to get the updated target model. In this paper, we present an incremental execution algorithm for ATL, as well as a prototype. With it, changes in a source model are immediately propagated to the target model. Our approach leverages previous works of the community, notably on live transformations and incremental OCL. We achieve our goal on a subset of ATL, without requiring modifications to the language.