Many existing clustering algorithms use a single prototype to represent a cluster. However sometimes it is very difficult to find a suitable prototype for representing a cluster with an arbitrary shape. One possible solution is to employ multi-prototype instead. In this paper, we propose a minimum spanning tree (MST) based multi-prototype clustering algorithm. It is a split and merge scheme. In the split stage, the suitable patterns are determined to be prototypes in terms of their degrees in the corresponding MST. Then the fake prototypes in sparse density regions are recognized and removed. In the merge stage, a two-step merge strategy is designed to group the prototypes. The proposed algorithm can deal with datasets consisting of clusters with different shapes, sizes and densities. The experiment results show the effectiveness on the synthetic and real datasets.