In AFNI, you would run 3dFWHMx on your residual time series (errts) to get estimates of smoothness for each individual. Assuming those estimates are similar across all subjects, you then average those smoothness values (xyz) as your input for 3dClustSim.
In SPM, you may be able to use spm_est_smoothness() on each individual subjects ResMS.hdr. The suggestion from Andy's Brian Blog [
andysbrainblog.blogspot.com]. There are some posts on the SPM mailing list on this topic suggesting similar courses of action.