The addition of macro and micro fibers can enhance the compressive strength of fiber-reinforced coral aggregate concrete (FRCAC-CS). Traditional explicit models for FRCAC-CS suffer from shortcomings related to accuracy and generalization capabilities. To address this issue, a comprehensive database for FRCAC-CS was established by collecting 834 sets of experimental data from 15 literature sources. The hunger games search algorithm (HGS), sooty tern optimization algorithm, and whales optimization algorithm were used to optimize the hyperparameters of the XGBoost algorithm, resulting in optimized XGBoost models for FRCAC-CS. The performance of the prediction models was assessed through comparisons between predicted and actual values, model residual distribution, and performance evaluation metrics. Sensitivity analysis was conducted using the Shapley additive explanations method. Subsequently, 9 different mix proportions were formulated to validate the performance of the optimal model. The results suggest that the HGS–XGBoost model yields predictions that are closer to the actual values, exhibiting smaller mean and standard deviation in model residuals. The five major performance evaluation metrics, including RMSE, MSE, MAPE, MAE, and R2, for the HGS–XGBoost model are 0.91, 0.83, 2.35, 0.77, and 0.99, respectively, surpassing those of other models. Sensitivity analysis highlights the water–to–binder ratio and total aggregate content as the two most crucial factors among the mix components. Decreasing the water–to–binder ratio and the total aggregate content improves the FRCAC-CS. The experimental results from the 9 mix proportions demonstrate that the error between predicted and experimental values is below 7%, affirming the HGS–XGBoost model’s adequate accuracy and generalization capabilities for new datasets. This study introduces a novel AI-based prediction method for FRCAC-CS, establishing the groundwork for its potential application in reef construction.