Although omnivory is common in nature, its impact on trophic interactions is variable. Predicting the food web consequences of omnivory is complicated because omnivores can simultaneously produce conflicting direct and indirect effects on the same species or trophic level. We conducted field and laboratory experiments testing the top-down impacts of an omnivorous salt marsh crab, Armases cinereum, on the shrub Iva frutescens and its herbivorous and predatory arthropod fauna. Armases is a "true omnivore," consuming both Iva and arthropods living on Iva. We hypothesized that Armases would benefit Iva through a top-down trophic cascade, and that this benefit would be stronger than the direct negative effect of Armases on Iva. A field experiment on Sapelo Island, Georgia (USA), supported this hypothesis. Although Armases suppressed predators (spiders), it also suppressed herbivores (aphids), and benefited Iva, increasing leaf number, and reducing the proportion of dead shoots. A one-month laboratory experiment, focusing on the most common species in the food web, also supported this hypothesis. Armases strongly suppressed aphids and consumed fewer Iva leaves if aphids were available as an alternate diet. Armases gained more body mass if they could feed on aphids as well as on Iva. Although Armases had a negative effect on Iva when aphids were not present, Armases benefited Iva if aphids were present, because Armases controlled aphid populations, releasing Iva from herbivory. Although Armases is an omnivore, it produced strong top-down forces and a trophic cascade because it fed preferentially on herbivores rather than plants when both were available. At the same time, the ability of Armases to subsist on a plant diet allows it to persist in the food web when animal food is not available. Because omnivores feed on multiple trophic levels, their effects on food webs may differ from those predicted by standard trophic models that assume that each species feeds only on a single trophic level. To better understand the complexity of real food webs, the variable feeding habits and feeding preferences of different omnivorous species must be taken into consideration.