Motivated by recent advances in Reinforcement Learning (RL) and the lack of open-source tools for training and benchmarking satellite guidance and control, we introduce LEO-GYM: a lightweight Python library for formulating RL problems for satellites in Low Earth Orbit (LEO). The framework decomposes problems into three classes, the low-level dynamics, the training environment and a satellite object that bridges them. LEO-GYM enables the creation of custom scenarios without imposing rigid class hierarchies. We present the architecture, key components, and an illustrative orbit-correction task modeled as a semi-Markov decision process. LEO-GYM is released as open-source to support and foster reproducible research in autonomous space operations.