Anomaly detection in multivariate time series (MTS) is critical for a wide range of applications. However, existing methods often inadequately capture coupled spatio-temporal dependencies and treat multiple temporal scales independently, overlooking cross-scale interactions that jointly determine anomalous behavior. This limitation often compromises the accuracy and robustness of anomaly detection. To address this issue, we propose MS-STNet, a multi-scale unified spatio-temporal modeling framework. Specifically, MS-STNet first decomposes the MTS into multiple temporal scales. At each scale, a Spatio-Temporal Encoder captures evolving spatio-temporal dependencies through the Temporal Context Mixer (TCMixer) and the Adaptive Channel Hypergraph Convolution (ACHConv). Furthermore, each scale serves as an expert, with a learnable routing mechanism that enables selective information exchange across scales, capturing complex inter-scale association. Extensive experiments on eight benchmark datasets demonstrate that MS-STNet achieves state-of-the-art performance in anomaly detection.