The majority of fish studies on coral reefs consider only non-cryptic species and, despite their functional importance, data on cryptic species are scarce. This study investigates inter-habitat variation in Caribbean cryptobenthic fishes by re-analysing a comprehensive data set from 58 rotenone stations around Buck Island, U.S. Virgin Islands. Boosted regression trees were used to associate the density and diversity of non-piscivorous cryptobenthic fishes, both in the entire data set and on reef habitats alone, with 14 abiotic and biotic variables. The study also models the habitat requirements of the three commonest species. Dead coral cover was the first or second most important variable in six of the eight models constructed. For example, within the entire data set, the number of species and total fish density increased approximately linearly with increasing dead coral cover. Dead coral was also important in multivariate analyses that discriminated 10 assemblages within the entire data set. On reef habitats, the number of species and total fish density increased dramatically when dead coral exceeded ~55 %. Live coral cover was typically less important for explaining variance in fish assemblages than dead coral, but live corals were important for maintaining high fish diversity. Coral species favoured by cryptobenthic species may be particularly susceptible to mortality, but dead coral may also provide abundant food and shelter for many fishes. Piscivore density was a key variable in the final models, but typically increased with increasing cryptobenthic fish diversity and abundance, suggesting both groups of fishes are responding to the same habitat variables. The density of territorial damselfishes reduced the number of cryptobenthic fish species on reef habitats. Finally, habitats delineated by standard remote sensing techniques supported distinct cryptobenthic fish assemblages, suggesting that such maps can be used as surrogates of general patterns of cryptic fish biodiversity in conservation planning.