Species inventories are the building blocks of our assessment of biodiversity patterns and human impact. Yet, historical inventories based on visual observations are often incomplete, impairing subsequent analyses of ecological mechanisms, extinction risk and management success. Environmental DNA (eDNA) metabarcoding is an emerging tool that can provide wider biodiversity assessments than classical visual-based surveys. However, eDNA-based inventories remain limited by sampling effort and reference database incompleteness. In this study, we propose a new framework coupling eDNA surveys and sampling-theory methods to estimate species richness in under-sampled and hyper-diverse regions where some species remain absent from the checklist or undetected by visual surveys. We applied this framework to the coastal fish diversity in the heart of the coral triangle, the richest marine biodiversity hotspot worldwide. Combining data from 279 underwater visual censuses, 92 eDNA samples and an extensive custom genetic reference database, we show that eDNA metabarcoding recorded 196 putative species not detected by underwater visual census including 37 species absent from the regional checklist. We provide an updated checklist of marine fishes in the ‘Raja Ampat Bird's Head Peninsula' ecoregion with 2534 species including 1761 confirmed and 773 highly probable presences. The Chao lower-bound diversity estimator, based on the incidence of rare species, shows that the region potentially hosts an additional 123 fish species, including pelagic, cryptobenthic and vulnerable species. The extended and hidden biodiversity along with their asymptotic estimates highlight the ability of eDNA to expand regional inventories and species distributions to better guide conservation strategies.