This chapter introduces the ZOOpt toolbox, a powerful tool for zeroth-order optimization designed to address high-dimensional and noisy optimization problems, particularly in machine learning tasks such as hyper-parameter tuning and direct policy search. ZOOpt implements state-of-the-art algorithms, including SRacos, ASRacos, and POSS, and supports optimization in continuous, discrete, and hybrid spaces. It also features noise-handling mechanisms like value suppression and threshold selection, as well as high-dimensionality handling through sequential random embedding. The toolbox integrates with the Ray framework for distributed optimization, enabling efficient parallel computation. Empirical studies demonstrate ZOOpt’s superior convergence rate, scalability, and robustness against noise compared to other optimization toolboxes. Experiments on synthetic functions and machine learning tasks, including classification with Ramp loss and OpenAI Gym control tasks, highlight ZOOpt’s effectiveness. The chapter concludes with a summary of ZOOpt’s capabilities and its potential for real-world applications.