Background: Membrane proteins are vastly underrepresented in structural databases, which has led to a lack of computational tools and the corresponding inappropriate use of tools designed for soluble proteins. For membrane proteins, lipid accessibility is an essential property. Even though programs are available for sequence-based prediction of lipid accessibility and structure-based identification of solvent-accessible surface area, the latter does not distinguish between water accessible and lipid accessible residues in membrane proteins. Results: Here we present mp_lipid_acc, the first method to identify lipid accessible residues from the protein structure, implemented in the RosettaMP framework and available as a webserver. Our method uses protein structures transformed in membrane coordinates, for instance from PDBTM or OPM databases, and a defined membrane thickness to classify lipid accessibility of residues. mp_lipid_acc is applicable to both -helical and -barrel membrane proteins of diverse architectures with or without water-filled pores and uses a concave hull algorithm for classification. We further provide a manually curated benchmark dataset, on which our method achieves prediction accuracies of 90%. Conclusion: We present a novel tool to classify lipid accessibility from the protein structure, which is applicable to proteins of diverse architectures and achieves prediction accuracies of 90% on a manually curated database. mp_lipid_acc is part of the Rosetta software suite, available at www.rosettacommons.org. The webserver is available at http://rosie.graylab.jhu.edu/mp_lipid_acc/submit and the benchmark dataset is available at http://tinyurl.com/mp-lipid-acc-dataset.