When prototyping story generation or conversational content within a company, there is sometimes a need to avoid sending the settings and character information entered into a cloud API. In validation environments where account registration and API key management should be avoided, or in situations involving highly confidential project concepts, whether the generation process itself can be confined to a local machine often becomes the deciding factor in selecting a tool. Open Dungeon is an MIT-licensed AI role-playing app published by newideas99. It is designed on the premise that both text generation and illustration generation can be completed locally.

Features Provided by Open Dungeon

According to the official README, text generation uses the Gemma 4 QAT model through Ollama by default, while the architecture can also connect to OpenAI-compatible backends such as llama.cpp, LM Studio, vLLM, a remote Ollama instance, and OpenRouter. Narration is said to be streamed in real time as the model writes it. For image generation, the narrator calls a tool named generate_image, causing an illustration suited to the scene to be generated inline. The README describes both a method for running FLUX.2-klein on Apple Silicon, NVIDIA, and compatible AMD Radeon hardware and a method for specifying your own ComfyUI instance. In terms of operation, it offers three input modes: Do for player actions, Say for dialogue, and Story for writing narrative prose directly, along with operations such as Continue, Retry, and Erase, as well as inline editing of individual utterances. For long stories, the conversation history is kept within the model's context window, which ranges from 128K to 256K, while older portions are compressed into summaries so that they are not forgotten. A character's saved portrait is used both as part of the narrator's visual context and as a reference image for image generation, helping maintain visual consistency. Data is stored in a local SQLite database and folders on disk, and the app can reportedly be played from a smartphone via Tailscale.

When It Is Worth Considering

Potential use cases include validation work with strict privacy requirements where generated stories and images must not be sent to external services, learning use cases where developers want to examine local LLM behavior and long-context management even on a machine without a GPU, and situations where a company already operates an OpenAI-compatible server and wants to use Open Dungeon as a reference example of a frontend running on top of it. The README lists multiple installation options, including a DMG for Mac, a batch launcher for Windows, and cloning the source code, making the ability to choose an approach based on the environment another point worth checking.

Points That Cannot Be Determined from the README Alone

The README does not provide specific figures for the quality of generated text and images, response speed, or required memory and storage capacity. The actual experience is thought to depend greatly on the environment and model selection. The supported range of AMD Radeon hardware, operational status in Linux environments, and other details of hardware and OS compatibility cannot be determined from the body of the article. In addition, the default narrator prompt is said to be designed to permit consensual adult content. When introducing it within an organization, note that this default behavior must be adjusted by modifying the system prompt.

This article summarizes the content described in the official README (https://github.com/newideas99/open-dungeon/blob/main/README.md) of newideas99/open-dungeon.

AI was used to assist with writing and editing based on the official README. This is not a report of verification on an actual device.