Several metamodels are often introduced when looking at the same fragment of the real world from different points of view. This process involves multiplying data to be instances of those metamodels thus enforcing the redundancy and complicating the maintenance. In this paper, a mechanism for defining (read only) views on arbitrary metamodels is presented. Having the mechanism, a user is allowed to work with the view metamodel and its instances by the means of a model transformation language while at the same time only instances of the base metamodel are kept physically in the memory. Thus, the proposed view definition mechanism does not introduce a redundancy of data. Also, the basic ideas of compiling model transformation programs from the view metamodel to the base metamodel are outlined here.