Accurate indoor crowd counting (ICC) is a key enabler to many smart home/office applications. Recent development of the WiFi-based ICC technology relies on detecting the variation of wireless channel state information (CSI) caused by human motions and has gained increasing popularity due to its low hardware cost, reliability under all lighting conditions, and privacy preservation in sensing data processing. To attain high estimation accuracy, existing WiFi-based ICC methods often require a large amount of labeled CSI training data samples for each application domain, i.e., a particular WiFi transceiver or background deployment. This makes large-scale deployment of the WiFi-based ICC technology across dissimilar domains extremely difficult and costly. In this article, we propose a Domain-Agnostic and Sample-Efficient wireless indoor crowd Counting (DASECount) framework that suffices to attain robust cross-domain detection accuracy given very limited data samples in new domains. DASECount leverages the wisdom of the few-shot learning (FSL) paradigm consisting of two major stages: 1) source domain meta training and 2) target domain meta testing. Specifically, in the meta-training stage, we design and train two separate convolutional neural network (CNN) modules on the source domain data set to fully capture the implicit amplitude and phase features of CSI measurements related to human activities. A subsequent knowledge distillation procedure is designed to iteratively update the CNN parameters for better generalization performance. In the meta-testing stage, we use the partial CNN modules to extract low-dimension features out of the high-dimension input target domain CSI data. With the obtained low-dimension CSI features, we can even use very few amounts of target domain data samples (e.g., 5-shot samples) to train a lightweight logistic regression (LR) classifier, and attain very high cross-domain ICC accuracy. Experiment results show that the proposed DASECount method achieves over 92.68%, and on average 96.37% detection accuracy in a 0-8 people counting task under various domain setups, which significantly outperforms the other representative benchmark methods considered.