TART is a package for training and deploying task-agnostic reasoning modules which can be composed with arbitrary foundation models to improve in-context learning peformance for classification tasks.