Towards Video Anomaly Detection (VAD), existing methods require labor-intensive data collection and model retraining, making them costly and domain-specific. The proposed method, termed as Multi-modal Caption Aware Network (MCANet), introduces a novel paradigm that identifies anomalies in video sequences without requiring prior domain knowledge. This training-free VAD approach dynamically generates and analyzes textual descriptions of video frames by utilizing off-the-shelf vision-language model (VLM), audio-language model (ALM) and large language model (LLM). MCANet has four primary modules. The first module utilizes image-text similarities to clean noisy captions generated by the image captioning model, while the second module applies audio-text similarities to refine noisy captions produced by the audio captioning model. The third module employs a LLM to consolidate scene dynamics over time. Finally, the fourth module enhances the results by aggregating scores from semantically similar frames based on video-text similarity. To validate the effectiveness of the proposed method, experiments are conducted on two large-scale benchmark datasets (UCF-Crime and XD-Violence). Experimental results demonstrate that MCANet surpasses existing unsupervised and one-class approaches without requiring any training or data collection.