Hi Mark,
I thought one of the main advantages for using the TLSmd server was that it provided the crystallographer with suggestions for the TLS staring groups (as well as the TLS values for these groups) avoiding the need to make somewhat arbitrary choices about the TLS group boundaries.
To do TLS refinement you need to define the selections for TLS groups. This can be done either manually by just looking at your model or by using TLSmd server. These selections for TLS groups is all phenix.refine needs to do TLS refinement. phenix.refine does not need the starting values for TLS, it finds them automatically from doing some preliminary group B-factors refinement and optimal splitting of total B into individual and group B (http://www.ccp4.ac.uk/newsletters/newsletter45/articles/Tmax-CCP4_Afonine-Ur...). Again, as far as I know, the TLSmd server automatically finds the selections for TLS groups using the B-factors from input PDB file. And I think this is a weak point about TLSmd since that B-factors can be biased by previously used restraints or somehow distorted, while phenix.refine uses the experimental data to find these starting values. Pavel.