OpenTinker is a platform that simplifies Agentic Reinforcement Learning. It allows users to set up environments for various learning tasks, ranging from single-turn mathematical problems to multi-turn interactions involving games like Gomoku and AlfWorld. The project provides step-by-step instructions for installation, training, and using different agents, making it accessible for developers and researchers. Key tasks include single-turn math problems, multi-turn interactions with tool usage, and geometry problem-solving, all of which can be monitored via the WandB (Weights & Biases) tool for performance tracking. To get started, users clone the OpenTinker repository and install dependencies, either through Docker or manually. The platform has a built-in authentication system to secure access to its API, requiring users to register and obtain an API key. The architecture is designed to support various training scenarios by accommodating two main dimensions: the data source and the interaction mode. This flexibility allows users to create environments that are either data-dependent or data-free, and to choose between single-turn or multi-turn interactions. OpenTinker’s design space results in four paradigms suited for different learning objectives: one-shot mathematical reasoning using structured datasets, iterative problem-solving with tool assistance, one-shot interactions from simulators, and complex game playing through iterative simulations. This structured approach enhances the learning experience by providing tailored environments that meet specific research or application needs. The project also emphasizes future improvements, aiming to decouple its client from the underlying training code to ensure a more lightweight and independent setup.