Accurately predicting the shear strength of jointed rock masses is crucial for intelligently evaluating the surrounding rock stability during tunnel construction. In this study, the authors propose a machine learning approach based on the XGBoost model, whose hyperparameters are optimized using metaheuristic algorithms. Five input features—joint roughness coefficient, uniaxial compressive strength, elastic modulus, normal stress, and shear surface length—are selected based on an analysis of joint shear behavior, with peak shear strength as the output. Correlation analysis identifies normal stress as the most significant factor influencing the shear strength. Three metaheuristic algorithms—equilibrium optimizer (EO), grey wolf optimizer (GWO), and slime mould algorithm (SMA)—are used to optimize the XGBoost model, and the optimal initial population size for each algorithm is determined. The performance of the resulting hybrid models (EO-XGBoost, GWO-XGBoost, and SMA-XGBoost) is compared with models optimized using random search, as well as traditional machine learning models such as random forest and support vector regression. The results demonstrate that all three metaheuristic-optimized models significantly outperform the baseline methods, with the SMA-XGBoost model achieving the highest predictive accuracy (RMSE=0.393 21, R²=0.996 24, MAE=0.256 89, and VAF=0.996 25). Furthermore, the Shapley additive explanation analysis reveals that NS is the most influential feature in the prediction process. Overall, this study provides a robust and efficient data-driven framework for accurately predicting joint shear strength, supporting real-time stability assessments in tunnel engineering.