The COVID-19 pandemic death toll now surpasses two million individuals and there is a need for early identification of individuals at increased risk of mortality. Host genetic variation partially drives the immune and biochemical responses to COVID-19 that lead to risk of mortality. We identify and prioritise blood proteins and biomarkers that may indicate increased risk for severe COVID-19, via a proteome Mendelian randomization approach by collecting genome-wide association study (GWAS) summary statistics for >4,000 blood proteins. After multiple testing correction, troponin I3, cardiac type (TNNI3) had the strongest effect (odds ratio (O.R.) of 6.86 per standard deviation increase in protein level), with proteinase 3 (PRTN3) (O.R.=2.48), major histocompatibility complex, class II, DQ alpha 2 (HLA-DQA2) (O.R.=2.29), the C4A-C4B heterodimer (O.R.=1.76) and low-density lipoprotein receptor-related protein associated protein 1 (LRPAP1) (O.R.=1.73) also being associated with higher odds of severe COVID-19. Conversely, major histocompatibility complex class I polypeptide-related sequence A (MHC1A) (O.R.=0.6) and natural cytotoxicity triggering receptor 3 (NCR3) (O.R.=0.46) were associated with lower odds. These proteins are involved in heart muscle contraction, natural killer and antigen presenting cells, and the major histocompatibility complex. Based on these findings, it may be possible to better predict which patients may develop severe COVID-19 and to design better treatments targeting the implicated mechanisms.