A three-dimensional variational data assimilation (3DVAR) system (ROMS3DVAR) has been developed for the Regional Ocean Modeling System (ROMS). This system provides a capability of predicting meso- to small-scale variations with temporal scales from hours to days in coastal oceans. Novel strategies are implemented, including the implementation of three-dimensional anisotropic and inhomogeneous error correlations, application of particular weak dynamic constraints, and implementation of efficient and reliable algorithms for minimizing the cost function. ROMS3DVAR has been implemented in a quasi-real-time fashion in support of both the Southern and Central California Coastal Ocean Observing System (SCCOOS and CenCOOS). ROMS3DVAR assimilates a variety of observations, including satellite sea surface temperatures and sea surface heights, High Frequency (HF) radar velocities, ship reports and other available temperature and salinity profiles. The evaluation of data assimilation and forecast skills is presented. Impacts of various observations, particularly HF radar velocities, are discussed.