OpenForgeRL is an open-source framework designed to enable end-to-end training of agents utilizing inference harnesses like Claude Code, Codex, and OpenClaw across diverse environments. It addresses the challenge of training stateful, multi-process harness-based agents using conventional SFT/RL infrastructure. By bridging the gap between complex harness systems and open training stacks, OpenForgeRL allows seamless integration of tool use and external system interactions during training. The framework is publicly available for community use and further development.