Purpose: A supervised learning framework is proposed to automatically generate MR sequences and corresponding reconstruction without human knowledge on MR strategies. This enables a target-based optimization from scratch, as well as exploration of novel and flexible MR sequence strategies. Methods: The entire scanning and reconstruction process is simulated end-to-end in terms of RF events, gradient moment events in x and y, and delay times, acting on the input model spin system given in terms of proton density, T1 and T2, and B0. As proof of concept we use both conventional MR images but also binary masks and T1 maps as a target and optimize from scratch using the loss defined by data fidelity, SAR, and scan time. Results: In a first attempt, MRzero learns all gradient and RF events from zero, and is able to generate the aimed at target image. Appending a neural network layer to the reconstruction module also arbitrary targets were learned successfully. Experiments could be translated to image acquisition at a real system (3T Siemens, PRISMA) and could be verified in measurements of phantoms and the human brain in vivo. Discussion/Conclusion: We have developed a fully automated MR sequence generator based on Bloch equation simulations and supervised learning. While we focus on the method herein, having such a differentiable digital MR twin at hand paves the way to a novel way of generating MR sequence and reconstruction solely governed by the target provided, which can be a certain MR contrast, but the possibilities for targets are limitless, e.g. quantification, segmentation, as well as contrasts of other image modalities.