The variables themselves are directly observed, however they are typically used to create one or more latent variables. For instance, if I have a 3-item depression measure, those three items would be observed, but they would create a single latent variable called "depression"
Broadly speaking, you can use (almost) any combination of latent and observed variables in SEM. The point is to set up a model that represents your assumptions (like a set of observed variables measuring an unobserved latent factor), and then test that model against the data.