langchain-ai/google-adk-realtime-deepagents-example

This repository is a small demo project that shows how to build a real-time voice agent you can talk to out loud. It combines Google's ADK framework, which runs a live back and forth voice conversation using Gemini's Live API, with a separate research tool built using LangChain's deepagents. You speak into your browser's microphone, the voice agent replies out loud, and you can interrupt it mid-sentence, since the system supports barge-in. When you ask a question that needs real research rather than a quick answer, the voice agent hands the question off to a deep research tool. That tool plans out several web searches using Tavily, a search service, gathers the results, and writes up a report, all in the background while the voice agent keeps the conversation going. Once the report is ready, the voice model reads it back to you conversationally instead of just reading raw text. Every research run can optionally be traced with LangSmith, a tool for inspecting what an AI agent did step by step, showing the plan, each web search, and the final report as one connected tree you can replay later. Under the hood, a FastAPI server bridges audio between the browser and Google's live API over a WebSocket connection, converting microphone audio into a format the model expects and streaming the model's spoken replies back to the browser. A small JavaScript frontend in the browser handles capturing the microphone and playing audio back smoothly. To run it, you need a Google AI Studio key and a Tavily key, both set in a local environment file, then you install dependencies and start the server. The project is explicitly a local demo: it has no login system, sessions are only kept in memory, and the author notes that authentication, encryption, and a proper session store would need to be added before running it somewhere public.