We present a real-time algorithm for rendering volumetric 3D Magic LensesTM having arbitrary convex shapes. During fragment processing the algorithm performs a second depth test using a shadow map. Exploiting the second depth test we are able to classify each fragment, with respect to its position relative to the lens volume. Using this classification we first render the geometry behind the lens volume, then the geometry intersecting the lens volume using a different visual appearance and finally the parts in front of the lens volume. Regardless of the shape of the lens volume just two additional rendering passes are needed. Furthermore there are no limitations to the choice of visual appearance used to enhance expressiveness of the virtual world. We will describe theoretical and practical aspects of the algorithm and our implementation, which is accelerated by current graphics hardware. Keywords 3D Magic Lenses, 3D Exploration, Toolglass, See-Through Tools