This paper presents a complete theory of credal networks, structures that associate convex sets of probability measures with directed acyclic graphs. Credal networks are graphical models for precise/imprecise beliefs. The main contribution of this work is a theory of credal networks that displays as much flexibility and representational power as the theory of standard Bayesian networks. Results in this paper show how to express judgements of irrelevance and independence, and how to compute inferences in credal networks. A credal network admits several extensions—several sets of probability measures comply with the constraints represented by a network. Two types of extensions are investigated. The properties of strong extensions are clarified through a new generalization of d-separation, and exact and approximate inference methods are described for strong extensions. Novel results are presented for natural extensions, and linear fractional programming methods are described for natural extensions. The paper also investigates credal networks that are defined globally through perturbations of a single network.