Peptide-based therapeutics are increasingly emerging as a promising alternative to small molecules and protein drugs. However, their clinical development still faces significant challenges, particularly in terms of efficacy and safety, which are largely attributed to their suboptimal absorption, distribution, metabolism, and excretion properties and potential toxicity risks (ADMET). To address these challenges, we developed pepADMET (https://pepadmet.ddai.tech), the first publicly accessible AI-driven platform for the systematic and comprehensive assessment of peptide ADMET properties. The platform integrated 36,643 high-quality data entries and covers 19 key ADMET endpoints. By combining molecular graph representations, enzymatic descriptors, and transfer learning with advanced neural architectures such as graph neural networks and relational graph convolutional networks, the platform effectively captures complex molecular and biological features of peptides, thereby substantially enhancing the predictive performance and robustness of the models. Specifically, we introduced MLR-GAT, a novel multilevel framework specifically designed for peptide toxicity prediction, which can hierarchically identify multiple categories of peptide toxicity rather than focusing solely on hemolytic toxicity. Uniquely, pepADMET for the first time simultaneously supports linear, cyclic, modified, and natural peptides, while also accounting for biological variability across species, organs, and cell lines, thereby enabling more precise and biologically relevant ADMET predictions. As a new comprehensive online resource for multiproperty ADMET evaluation of peptides, pepADMET provides a unified, accurate, and intelligent framework to advance peptide drug design and development.