Purpose <p>This work presents a system based on machine learning for the acquisition and processing of sEMG signals for online applications in hand gesture recognition using the Myo armband. This study aims to present a real-time classification tool, exploring processing parameters to optimize the online application, which is the novelty of this work.</p> Methods <p>Five hand gestures from five subjects were acquired by the Myo armband and recognized in this study. Offline analyses were employed to select the best parameters for online operations, including segmentation, feature extraction, and classification. The algorithms’ training strategies are compared individually and with all volunteers. Our proposed algorithm for muscle activation detection presented higher performance and lower latency than the double-threshold onset method. The classification models trained were Decision Tree, Bagging, Random Forest, Support Vector Machine (SVM), and Linear Discriminant Analysis.</p> Results <p>During offline training, the use of Hudgins’ Feature Set, windows of 1.25 s, and an SVM classifier reached the best accuracy. In online operation, it represented an accuracy of 77.6% when considering data from all volunteers for training. Accuracy improved to 93.6% when training and testing steps in online operation were conducted separately for volunteers.</p> Conclusion <p>The online gesture classification tool was developed, exploring the parameters and process used for an online system. Every step for a gesture recognition system using sEMG was analyzed, from data collection to final classification, with a new proposed algorithm for online muscle activation segmentation.</p>