WiFi-based sensing technology has become a popular research direction in the Internet of Things (IoT) field. However, the accuracy of action sensing across different environmental domains is severely degraded when the model is deployed at the real IoT edge. Existing cross-domain sensing methods are unsuitable for real-life applications due to the large amount of expensive channel state information (CSI) data required for training. Meanwhile, pretrained predictive models in the cloud may not perform well in edge-side deployment environments. To address these issues, we propose a few-shot cross-domain WiFi sensing (FewCS) system with online learning. The model aggregates unlabeled samples from the same target domain and separates samples from different domains while minimizing sample cost and training an accurate WiFi-sensing system. The core idea of FewCS is to capture the sensing features of data through intradomain prototype clustering and perform cross-domain prototype extraction in a shared embedding space of multiple data. Moreover, we extend the model to the IoT edge segment for online learning and fine-tune the model parameters to fit the new domain on the cloud by using data collected in the field. Extensive experimental results on real datasets show that the proposed scheme significantly outperforms the current state-of-the-art WiFi sensing methods in terms of sensing accuracy and achieves satisfactory online learning performance with fewer training samples and times.