Neural weather models have shown immense po-tential as inexpensive and accurate alternativesto physics-based models. However, most mod-els trained to perform weather forecasting do notquantify the uncertainty associated with their fore-casts. This limits the trust in the model and theusefulness of the forecasts. In this work we con-struct and formalise a conformal prediction frame-work as a post-processing method for estimatingthis uncertainty. The method is model-agnosticand gives calibrated error bounds for all variables,lead times and spatial locations. No modifica-tions are required to the model and the computa-tional cost is negligible compared to model train-ing. We demonstrate the usefulness of the con-formal prediction framework on a limited areaneural weather model for the Nordic region. Wefurther explore the advantages of the frameworkfor deterministic and probabilistic models.