CompNav · Open-source projects
Open Dungeon
A local AI interactive-fiction app that generates roleplay stories and in-scene images.
What it is
Open Dungeon turns AI roleplay into a single-player interactive-fiction app. Users advance the story through actions, dialogue, or their own narration while an AI narrator describes scenes and characters and can generate illustrations in context. Story data is stored locally, and long histories are summarized for continuation. The project focuses on fictional worlds and plots rather than everyday companionship.
Core mechanisms and capabilities
Story input distinguishes player actions, character dialogue, and user-written narration. Users can continue, retry, edit, or erase passages, and character portraits can guide visual continuity in generated images. Older events are summarized to help long stories continue after the model context window fills.
Editorial take
It merits inclusion because it combines local models, interactive fiction, long-story memory, and scene images in a complete desktop/web app, while documenting model options and local data storage. The author's default text path is described as requiring no account or API key; remote providers or image generation change the dependencies and data flow.
Who it suits
Suited to people who enjoy solo text adventures, interactive fiction, or AI roleplay and prefer their stories and chats to remain local.
Deployment and technical requirements
Users can run a desktop release or launch the web app locally. The Windows launcher checks for Node.js and starts the service; Ollama and image models are optional. Official guides separate text and image setup so users who want text-only stories can skip the image environment.
Runtime requirements
The project provides a release for Apple Silicon Macs; on Windows, its launcher installs Node.js dependencies and starts the local app. Ollama is the default text-model route, with options to connect LM Studio, llama.cpp, vLLM, or another OpenAI-compatible backend. Image generation is optional and requires a separate image service.
The documented text path can use a local Gemma 4 QAT model through Ollama or another OpenAI-compatible backend; the default flow is described as not requiring an API key. Image models or ComfyUI are optional and require separate services, adding hardware or deployment needs.
Limitations and considerations
Output quality and narrative coherence depend on the connected model. Local inference reduces data sent to outside services but requires model downloads and local resources. Image generation uses a separate service, and the Mac release may require manually approving an unsigned app on first launch.
Official sources and further reading
Information checked: Oct 26, 2026.