We propose a self-training approach that enables large language models to learn tool use without requiring human demonstrations. Our method uses self-generated examples and iterative refinement to develop tool-use capabilities, significantly reducing the annotation burden while maintaining competitive performance on tool-use benchmarks.