"Fully Responsive Intelligent Digital Assistant for You"
Almost 1,000 stars on GitHub!
I developed and open-sourced F.R.I.D.A.Y. to bridge the gap between conversational AI and functional backend execution. Since its release, the repository has resonated strongly with the developer community, crossing 650 stars on GitHub.
The Problem
Building a voice assistant that is both highly responsive and deeply integrated with system tools is challenging. Traditional agents often bundle the voice pipeline and the tool-execution logic together, resulting in monolithic codebases that are hard to scale, suffer from high latency, or break conversational flow when executing complex external tasks.
The Solution
F.R.I.D.A.Y. splits the architecture into two cooperating layers: a lightning-fast voice pipeline that listens and speaks, and a robust backend server that acts as the "brain," doing the actual heavy lifting.
Key Features
- 🎙️ Real-Time Voice Pipeline: Powered by LiveKit Agents, providing seamless, interruptible voice interactions.
- 🧠 Decoupled Brain (MCP Server): A FastMCP server exposing tools (news, web search, system info) over Server-Sent Events (SSE).
- ⚡ Multi-Model Intelligence: Intelligently routes tasks combining Sarvam Saaras v3 for optimized STT, Gemini 2.5 Flash for rapid reasoning, and OpenAI's Nova voice for high-fidelity TTS.
- 🛠️ Extensible Architecture: New capabilities can be added to the MCP backend without ever touching the core voice agent loop.
Tech Stack
The architecture relies on modern, high-performance tooling:
- Frameworks: FastMCP, LiveKit Agents
- AI / Models: Google Gemini 2.5 Flash (LLM), Sarvam Saaras v3 (STT), OpenAI (TTS)
- Language & Package Management: Python 3.11+,
uv