// the find
lss233/kirara-ai
🤖 可 DIY 的 多模态 AI 聊天机器人 | 🚀 快速接入 微信、 QQ、Telegram、等聊天平台 | 🦈支持DeepSeek、Grok、Claude、Ollama、Gemini、OpenAI | 工作流系统、网页搜索、AI画图、人设调教、虚拟女仆、语音对话 |
Kirara AI is a Python chatbot framework for connecting LLMs (OpenAI, Claude, DeepSeek, Gemini, etc.) to messaging platforms like WeChat, QQ, and Telegram. It ships with a visual workflow editor, plugin marketplace, and web admin UI. The target audience is Chinese-market self-hosters who want a configurable AI assistant on platforms that Western tools ignore.
The workflow system is genuinely well-architected — blocks have typed inputs/outputs, there's a registry, a dispatcher, and YAML-serializable workflow definitions, which is more disciplined than most bots in this space. The LLM adapter layer is clean: each provider gets its own adapter file with a shared request/response format, making it easy to add new models. The plugin architecture is real — plugins are proper packages with their own setup.py, not just monkey-patching. Test coverage exists for adapters, memory, and system blocks, with mock servers for Gemini/Ollama/OpenAI, which is more than most projects this size bother with.
Discord and Feishu support is listed as 'under reconstruction' — if those are your platforms, this is not ready. The Chinese-first documentation means most configuration, troubleshooting, and community help requires reading Chinese; the English README is a thin wrapper over the Chinese one. The bundled frpc (intranet tunneling) is a notable attack surface — it downloads and runs a binary to expose local ports, and there's no obvious audit trail or signature verification in the code. The web auth system lives in kirara_ai/web/auth/ but it's not clear what the threat model is; running this exposed to the internet without understanding those auth routes is risky.