Nasal endoscopy captures highly consistent regions of anatomical structures within the nasal cavity but also reveals fluid anomalies such as clear nasal discharge (CND) and purulent secretion (PUS), which are difficult to detect due to their transparency, low contrast, and specular reflection. To address these challenges, we propose a multi-level nasal endoscopy visual-language adapter (MEN-AD) for zero-shot and few-shot mucous anomaly detection. MEN-AD is based on a visual-language architecture, integrating a domain-adaptive feature enhancement module, a clinically inspired prompt library, and a structure-aware continuity loss function to achieve structure-preserving localization. We frame anomaly detection as a multi-modal alignment and dense classification task. MEN-AD adapts visual-language representations, enhancing the perception of ambiguous fluid regions and achieving high-accuracy mucous type recognition with minimal supervision. Additionally, we construct two internally annotated clinical datasets—CND and PUS—covering over 500 nasal endoscopy video sequences for systematic evaluation of this task. Extensive experimental results show that MEN-AD consistently outperforms the latest MVFA method in few-shot settings, especially in the 2-shot and 4-shot scenarios of the PUS dataset, achieving the highest improvements of +2.63% in classification accuracy and +5.07% in cosine similarity, demonstrating MEN-AD’s strong generalization ability under specular interference and weak boundary anomalies. To our knowledge, MEN-AD is the first vision-language framework for endoscopic fluid anomaly detection, offering a universal, annotation-efficient solution for ENT diagnosis.