Despite a rapid expansion in the number of known RNA viruses following the advent of metagenomic sequencing, the identification and annotation of highly divergent RNA viruses remains challenging, particularly from poorly characterized hosts and environmental samples. Protein structures are more conserved than primary sequence data, such that structure-based comparisons provide an opportunity to reveal the viral “dusk matter”: viral sequences with low, but detectable, levels of sequence identity to known viruses with available protein structures. Here, we present a new open computational and resource – RdRp-scan – that contains a standardized bioinformatic toolkit to identify and annotate divergent RNA viruses in metagenomic sequence data based on the detection of RNA dependent RNA polymerase (RdRp) sequences. By combining RdRp-specific Hidden Markov models (HMM) and structural comparisons we show that RdRp-scan can efficiently detect RdRp sequences with identity levels as low as 10% to those from known viruses and not identifiable using standard sequence-to-sequence comparisons. In addition, to facilitate the annotation and placement of newly detected and divergent virus-like sequences into the known diversity of RNA viruses, RdRp-scan provides new custom and curated databases of viral RdRp sequences and core motif, as well as pre-built RdRp alignments. In parallel, our analysis of the sequence diversity detected by RdRp-scan revealed that while most of the taxonomically unassigned RdRps fell into pre-established clusters, some sequences cluster into potential new orders of RNA viruses related to the Wolframvirales and Tolivirales. Finally, a survey of the conserved A, B and C RdRp motifs within the RdRp-scan sequence database revealed additional variations of both sequence and position, which might provide new insights into the structure, function and evolution of viral RdRps.