NitroGen is an open foundation model designed for generalist gaming agents, capable of interpreting pixel inputs and predicting gamepad actions. Developed through behavior cloning, it utilizes the largest video-action gameplay dataset sourced from internet videos. While it has potential for adaptation to new games after training, the current version is limited; it consists of a 500 million parameter DiT model that can only process the last frame of gameplay. This restricts its ability to plan over extended periods, play games continuously, or improve itself over time. To set up NitroGen, users need to clone the GitHub repository and install the necessary package with Python 3.12 or higher. The model requires a Windows environment to run games, although it can be served from a Linux machine for inference. Users must locate the exact executable name of the game they wish to play, which can be found in the Windows Task Manager. The instructions for running the agent are straightforward, involving a couple of Python commands to start the inference server and launch the game. The project is intended strictly for research, not as a commercial product from NVIDIA. For those interested in the underlying research, a citation for the relevant paper is provided, detailing the work of multiple contributors. This model represents a significant step in developing versatile gaming agents, but current limitations on planning and game interaction highlight areas for future improvement.