The determination of reference and referential links in discourse is one of the important challenges in natural language understanding. The first commonly adopted step towards this objective is to determine coreference classes over the set of referring expressions. We present a modular framework for automatic anaphora resolution which makes it possible to specify various anaphora resolution algorithms and to use them to build AR systems, in principle, for any natural language. The functionality of the system is shown on selected salience-based algorithms customized for Czech.