Wi-Fi sensing-based human activity recognition (HAR) research has grown over the last decade. While conventional Wi-Fi sensing employs complex and high-cost devices, we focused on lightweight Wi-Fi sensing with ESP32 for channel state information (CSI) data collection. We aim to keep the setup minimal by considering a single antenna for relatively static activities, e.g., sitting, standing, and light walking. Thus, we leverage CSI amplitude and propose EnSta-Fi, a classification model based on ensemble stacking for Wi-Fi sensing, combining baseline machine learning models, i.e., k-nearest neighbor (kNN) and support vector machine (SVM), with the logistic regression as a final classifier. Our method includes the actual measurement setup to collect CSI, ensemble stacking model training, and evaluation. Results showed that EnSta-Fi outperforms individual kNN and SVM in terms of activity classification performance with accuracy improvements of 2.29 and 1.19%, respectively. Moreover, compared with deep learning models, e.g., bidirectional gated recurrent unit (Bi-GRU) and convolutional neural network (CNN), EnSta-Fi achieves higher accuracy and less computational time (40 and 2.5 times faster than Bi-GRU and CNN, respectively). From the results of our proposed method, we can conclude that EnSta-Fi is suitable for the real deployment of the HAR system, where straightforward setup, light weight, high accuracy, and low computational complexity are emphasized.