The extraction of metadata used during the planning phase in mediation systems assumes the existence of a metadata repository that in most cases must be created with high human involvement. This dependency rises complexity of maintenance of the system and therefore the reliability of the metadata itself. This article presents MetaExtractor, a system which extracts structure, quality, capability and content metadata of structured data sources available on a mediation system. MetaExtractor is designed as a Multi-Agent System(MAS) where each agent specializes in the extraction of a particular type of metadata. The MAS cooperation capability allows the creation and maintenance of the metadata repository. MetaExtractor is useful to reduce the number of data sources selected during query planning in large scale mediation systems due to its ability to prioritize data sources that better contribute to answer a query. The work reported in this paper presents the general architecture of MetaExtractor and emphasizes on the extraction logic of content metadata and the strategy used to prioritize data sources accordingly to a given query.