What is this?
An open-source Python agent framework that connects a digital human or an LLM to your business systems: it adapts to various avatar model tech and LLMs on top, and serves digital-human APIs down to microcontrollers, apps and websites.
An open-source Python agent framework that connects a digital human or an LLM to your business systems: it adapts to various avatar model tech and LLMs on top, and serves digital-human APIs down to microcontrollers, apps and websites.Do I run it as a service, or import it as a library?You run it as a service — it supports both server mode and standalone mode — and your clients talk to it through text, voice, digital-human-driving, admin-control, auto-broadcast and intent interfaces.
You run it as a service — it supports both server mode and standalone mode — and your clients talk to it through text, voice, digital-human-driving, admin-control, auto-broadcast and intent interfaces.How do I get it running?Run it from source on Python 3.12 (Windows, macOS or Ubuntu — on Ubuntu install gcc and portaudio first): install the dependencies, rename system.conf.bak to system.conf and fill in your settings, then open the admin page in a browser.
Run it from source on Python 3.12 (Windows, macOS or Ubuntu — on Ubuntu install gcc and portaudio first): install the dependencies, rename system.conf.bak to system.conf and fill in your settings, then open the admin page in a browser.Which LLMs and voice models can I plug in?Everything is freely swappable: any OpenAI-compatible LLM endpoint, including DeepSeek and its thinking models, plus your choice of ASR and TTS engines — FunASR supplies speech recognition.
Everything is freely swappable: any OpenAI-compatible LLM endpoint, including DeepSeek and its thinking models, plus your choice of ASR and TTS engines — FunASR supplies speech recognition.What kind of clients can connect to it?Almost anything: microcontrollers, mobile apps, websites, big screens and third-party business systems, with multi-user multi-channel concurrency and a browser-based admin page for management.
Almost anything: microcontrollers, mobile apps, websites, big screens and third-party business systems, with multi-user multi-channel concurrency and a browser-based admin page for management.Can it run fully offline?Yes — fully offline operation is a listed feature, so with local ASR, TTS and LLM choices the whole voice-and-chat loop can stay on your own hardware.
Yes — fully offline operation is a listed feature, so with local ASR, TTS and LLM choices the whole voice-and-chat loop can stay on your own hardware.What kind of avatars can it drive?2.5D and 3D digital humans — including UE5 and Unity projects per the repo topics — on mobile, PC or web, with facial-expression output supported. Hooking up a digital human is optional: it also runs as a plain LLM or voice agent.
2.5D and 3D digital humans — including UE5 and Unity projects per the repo topics — on mobile, PC or web, with facial-expression output supported. Hooking up a digital human is optional: it also runs as a plain LLM or voice agent.Can it start a conversation on its own, or only respond?It can talk first: there is an auto-broadcast mode for virtual teachers, virtual streamers and news anchors, schedule-based proactive dialogue, plus wake-word activation with interruption support.
It can talk first: there is an auto-broadcast mode for virtual teachers, virtual streamers and news anchors, schedule-based proactive dialogue, plus wake-word activation with interruption support.Can I give it my own knowledge base and personality?Yes — custom knowledge bases, custom Q&A pairs and custom persona info are supported, and it keeps bionic memory so the character stays consistent. Voice commands are configured in qa.csv.
Yes — custom knowledge bases, custom Q&A pairs and custom persona info are supported, and it keeps bionic memory so the character stays consistent. Voice commands are configured in qa.csv.Does it support tool calling and MCP?Yes — the agent can decide on its own when to call tools, and Fay includes MCP tool management with SSE and studio modes. The credits note the tool-calling design references OpenAI Codex and the memory and skills design references openclaw.
Yes — the agent can decide on its own when to call tools, and Fay includes MCP tool management with SSE and studio modes. The credits note the tool-calling design references OpenAI Codex and the memory and skills design references openclaw.Can I use it in a commercial product?The code is GPL-3.0 and the README describes the project as fully open source with a commercial-use disclaimer; for business integrations the author lists a QQ contact.
The code is GPL-3.0 and the README describes the project as fully open source with a commercial-use disclaimer; for business integrations the author lists a QQ contact.What are similar projects?agentscope, TalkLikeYou-Streaming-RealTime-Demo, gpt4free, leon, aichat, agenticSeek, unsloth, V3SP3R
agentscope, TalkLikeYou-Streaming-RealTime-Demo, gpt4free, leon, aichat, agenticSeek, unsloth, V3SP3RWhich topics is it in?digital-human, unity3d, unreal-engine, asr, voice-assistant, tts, agent-framework, openai-compatible, chatbot, offline, deepseek, gplv3
digital-human, unity3d, unreal-engine, asr, voice-assistant, tts, agent-framework, openai-compatible, chatbot, offline, deepseek, gplv3