Background: The metabolome of COVID-19 patients has been studied sparsely, with most research focusing on a limited number of plasma metabolites or small cohorts. This is the first study to test saliva metabolites in COVID-19 patients in a comprehensive way, revealing patterns significantly linked to disease and severity, highlighting saliva’s potential as a non-invasive tool for pathogenesis or diagnostic studies. Methods: We included 30 asymptomatic subjects with no prior COVID-19 infection or vaccination, 102 patients with mild SARS-CoV-2 infection, and 61 hospitalized patients with confirmed SARS-CoV-2 status. Saliva samples were analyzed using hydrophilic interaction liquid chromatography–mass spectrometry (HILIC-MS/MS) in positive and negative ionization modes. Results: Significant differences in metabolites were identified in COVID-19 patients, with distinct patterns associated with disease severity. Dipeptides such as Val-Glu and Met-Gln were highly elevated in moderate cases, suggesting specific protease activity related to SARS-CoV-2. Acetylated amino acids like N-acetylserine and N-acetylhistidine increased in severe cases. Bacterial metabolites, including muramic acid and indole-3-carboxaldehyde, were higher in mild–moderate cases, indicating that oral microbiota differs according to disease severity. In severe cases, polyamines and organ-damage-related metabolites, such as N-acetylspermine and 3-methylcytidine, were significantly increased. Interestingly, most metabolites that were reduced in moderate cases were elevated in severe cases. Conclusions: Saliva metabolomics offers insightful information that is potentially useful in studying COVID-19 severity and for diagnosis.