open-dungeon

Open Dungeon is an easy-to-use, fully local AI roleplay application designed for private, on-device fiction. It generates narrative text and inline scene images entirely on your own machine, eliminating the need for accounts, API keys, or cloud services. The text generation is powered by Ollama using the Gemma 4 QAT model, or it can connect to any OpenAI-compatible backend such as llama.cpp, LM Studio, or vLLM. For visual content, the app uses local image generation with FLUX.2-klein on Apple Silicon, NVIDIA, or supported AMD GPUs, or it can integrate with a personal ComfyUI instance. It features full play controls including Do, Say, and Story inputs, along with options to Continue, Retry, Erase, or Edit passages. The application supports long-story memory by filling the model's context window and compacting older passages into a rolling summary. Characters maintain visual continuity through saved portraits that feed into both the narrator's vision context and the image generator. All data is stored privately in a local SQLite database and folders on your disk, allowing you to play from your phone over Tailscale. The project is MIT-licensed and provides quick-start guides for Mac, Windows, and cloning from source.