This paper studies secrecy capacity in a semi-deterministic setting, in which the channel between legitimate users (called Alice and Bob) is deterministic, while that between Alice and the eavesdropper (called Eve) is a discrete memoryless channel. Such a model is particularly relevant when a pre-existing error correcting code tailored to the legitimate channel is in use on top of which secret information is to be shared. First, a point-to-point setting is considered with a single wire-tapper, a situation in which the secrecy capacity has an elegant characterization. Next, a generalized multiple access setting with confidential messages is considered in which each user wishes to communicate secret information to a common destination without the other determining its message. In this latter situation, outer bounds on the secrecy capacity are obtained.