We present Toribash Learning Environment (ToriLLE), an interface with video game Toribash for training machine learning agents. Toribash is a MuJoCo-like environment of two humanoid character fighting each other hand-to-hand, controlled by changing states of body joints. Competitive nature of Toribash lends itself to two-agent experiments, and active player-base can be used for human baselines. This white paper describes the environment with its pros, cons and limitations as well experimentally show ToriLLE's applicability as a learning environment by successfully training reinforcement learning agents that improved over time. The code is available at https://github.com/Miffyli/ToriLLE.