Crowd counting has seen significant advancements with Convolutional Neural Networks (CNNs) and Transformers. CNNs excel in extracting local features, whereas Transformers leverage self-attention for global feature representation. However, CNNs struggle with capturing long-range dependencies, and Transformers incur substantial computational costs. Recently, Mamba, a state-space model, has demonstrated superior capability in capturing long-term dependencies with linear computational complexity. To exploit Mamba’s potential, we propose MCAttention, a plug-and-play module that enhances CNN-based models. Specifically, we develop a generalized Mamba cross-attention mechanism, which serves as an independent feature extraction branch, integrating global information while preserving CNNs’ local feature strengths. Rigorous experimental validations affirm the superiority of our proposed method, showcasing its compelling efficacy across diverse benchmarks. To promote transparency and facilitate future research endeavors, we will make the implementation code publicly accessible.