Based on a diverse and complementary set of theoretical and empirical findings, we describe an approach to phonology in which sound patterns are shaped by the trade-off between biases supporting message transmission accuracy and resource cost. We refer to this approach as Message-Oriented Phonology. The evidence suggests that these biases influence the form of messages, defined with reference to a language's morphemes, words or higher levels of meaning, rather than influencing phonological categories directly. Integrating concepts from information theory and Bayesian inference with the existing body of phonological research, we propose a testable model of phonology that makes quantitative predictions. Moreover, we show that approaching language as a system of message transfer provides greater explanatory coverage of a diverse range of sound patterns.