This paper presents a new tool, called Xfhl, for the automatic generation of hierarchical fuzzy systems. The tool takes a training dataset as input and finds the best hierarchical decomposition in terms of fuzzy modules with two inputs and one output. The behavior of Xfhl is based on an exhaustive search among all possible module decompositions. In order to evaluate each structure, Xfhl uses the mean square error on the structure after a process of parametric optimization. The tool is integrated into Xfuzzy 3, a development environment for fuzzy systems. © 2010 IEEE.