Almost all triangle mesh compression algorithms to date are driven by the mesh connectivity code. The geometry code usually employs a straightforward prediction method applied to the vertex sequence as dictated by the connectivity code. This generates a suboptimal geometry code, which results in significant loss in code efficiency, since the geometry dominates the mesh information content. This paper proposes a manifold mesh code which optimizes the geometric component, at the slight expense of the connectivity code. This mesh geometry code is shown to be up to 50% more compact than the state-of-the-art geometry code of Touma and Gotsman, especially for models with non-smooth geometry, such as CAD models.