Tuberculosis (TB) stands as the second most fatal infectious disease after COVID‐19, the effective treatment of which depends on accurate diagnosis and phenotyping. Metabolomics provides valuable insights into the identification of differential metabolites for disease diagnosis and phenotyping. However, TB diagnosis and phenotyping remain great challenges due to the lack of a satisfactory metabolic approach. Here, a metabolomics‐based diagnostic method for rapid TB detection is reported. Serum metabolic fingerprints are examined via an automated nanoparticle‐enhanced laser desorption/ionization mass spectrometry platform outstanding by its rapid detection speed (measured in seconds), minimal sample consumption (in nanoliters), and cost‐effectiveness (approximately $3). A panel of 14 m z−1 features is identified as biomarkers for TB diagnosis and a panel of 4 m z−1 features for TB phenotyping. Based on the acquired biomarkers, TB metabolic models are constructed through advanced machine learning algorithms. The robust metabolic model yields a 97.8% (95% confidence interval (CI), 0.964‐0.986) area under the curve (AUC) in TB diagnosis and an 85.7% (95% CI, 0.806‐0.891) AUC in phenotyping. In this study, serum metabolic biomarker panels are revealed and develop an accurate metabolic tool with desirable diagnostic performance for TB diagnosis and phenotyping, which may expedite the effective implementation of the end‐TB strategy.