Functional connectivity studies conducted at the group level using magnetoencephalography (MEG) suggest that resting state networks (RSNs) emerge from the large-scale envelope correlation structure within spontaneous oscillatory brain activity. However, little is known about the consistency of MEG RSNs at the individual level. This paper investigates the inter- and intra-subject variability of three MEG RSNs (sensorimotor, auditory and visual) using seed-based source space envelope correlation analysis applied to 5 min of resting state MEG data acquired from a 306-channel whole-scalp neuromagnetometer (Elekta Oy, Helsinki, Finland) and source projected with minimum norm estimation. The main finding is that these three MEG RSNs exhibit substantial variability at the single-subject level across and within individuals, which depends on the RSN type, but can be reduced after averaging over subjects or sessions. Over- and under-estimations of true RSNs variability are respectively obtained using template seeds, which are potentially mislocated due to inter-subject variations, and a seed optimization method minimizing variability. In particular, bounds on the minimal number of subjects or sessions required to obtain highly consistent between- or within-subject averages of MEG RSNs are derived. Furthermore, MEG RSN topography positively correlates with their mean connectivity at the inter-subject level. These results indicate that MEG RSNs associated with primary cortices can be robustly extracted from seed-based envelope correlation and adequate averaging. MEG thus appears to be a valid technique to compare RSNs across subjects or conditions, at least when using the current methods.