An automated refinement procedure (ARP) for protein models is proposed, and its convergence properties discussed. It is comparable to the iterative least-squares minimization/difference Fourier synthesis approach for small molecules. ARP has been successfully applied to three proteins, and for two of them resulted in models very similar to those obtained by conventional least-squares refinement and rebuilding with FRODO. In real time ARP is about ten times faster than conventional refinement. In its present form ARP requires high (2.0 A or better) resolution data, which should be of high quality and a starting protein model having about 75% of the atoms in roughly the correct position. For the third protein at 2.4 A resolution, ARP was significantly less powerful but nevertheless gave definite improvement, in the density map at least.