An operational system for production of site-specific probability forecasts from the ECMWF Ensemble Prediction System (EPS) will be described. A Kalman Filter (KF) is used to derive surface temperature, wind speed and precipitation from the model fields of each ensemble member, interpolated to local sites. As well as removing local biases, the KF also allows maximum and minimum temperatures to be derived statistically from standard model output. The methods used to develop a suitable KF for use with ensemble data will be described. Ensemble probabilities may then be calibrated based on verification rank histograms. Tails of the forecast distribution are modelled with parametric functions (Weibull distributions) fitted to observations lying outside the ensemble distribution in past verification. Methods of generating suitable calibration statistics will be described. Verification results will be shown demonstrating that in most cases both the KF and the calibration contribute to improving the probability forecasts, resulting in an operational system capable of generating high-quality probability forecasts. However, for some more severe weather events the calibration, which is based on statistics dominated by non-severe conditions, can actually degrade the ensemble probabilities.