CompNav · Open-source projects
Warashi
A Live2D desktop AI companion with long-term memory, proactive conversation, and interruptible voice chat.
What it is
Warashi turns a Live2D character into a desktop AI companion. It can remember information, start a topic after a quiet period, support voice barge-in, and use sleep or do-not-disturb mode; it also includes character management. Built on Open-LLM-VTuber, Warashi adds a first-run wizard, memory, and character controls. Users launch the local service and interact through a browser.
Core mechanisms and capabilities
Each character can have its own memory, personality, Live2D avatar, and voice. Long-term memory can be capped and consolidated. Voice chat supports interruption, the companion can start a topic during quiet periods, and saying goodnight activates do-not-disturb mode.
Editorial take
Beyond upstream voice and avatar capabilities, Warashi packages memory, proactive interaction, and first-run setup into a desktop-oriented companion experience with its own product focus. Its source and compiled frontend have different terms, while Live2D samples carry separate restrictions. Recent repository activity includes documentation and translations; the page reports maintenance based on repository state.
Who it suits
Suited to users who want an on-screen character with persistent memory and voice interaction and can provide a model locally or through an API.
Deployment and technical requirements
The project provides launch scripts for Windows and macOS. Users extract the release, run the launcher, let it install dependencies, and configure an LLM in the browser; source-based startup with Python and uv is also documented. The release avoids some manual steps but still requires an AI model.
Runtime requirements
Users choose a local Ollama model or hosted LLM API. First launch installs a Python environment and dependencies and downloads an approximately 1 GB speech-recognition model. Live2D characters and voice assets have their own terms, which should be checked before commercial use.
It can connect to OpenAI, Claude, or Gemini APIs, or local Ollama. The project gives qwen2.5:3b as a local default example and describes hosted APIs, Ollama Cloud, and other free or paid services. Model cost, privacy, and quality depend on the selected provider.
Limitations and considerations
Warashi is a friendlier package built on Open-LLM-VTuber, so upstream components and frontend terms still apply. Warashi's additions are MIT, but the full release also includes a frontend with additional conditions and Live2D samples. Hosted-model data is handled by the chosen provider; local models require adequate memory and inference resources.
Official sources and further reading
Information checked: Oct 26, 2026.