gemini-live-bootstrap

gemini-live-bootstrap

Set up an ADK 2.0 agent on the Gemini Live API for real-time streaming responses (text, audio, vision).

When to use

• Voice assistants where users expect immediate audio response
• Live transcription / translation flows
• Real-time multimodal experiences (camera + voice)
• Replacing turn-based chat with continuous interaction

Server-side template

# server.py
from google.adk.agents import LlmAgent
from google.adk.runners import Runner
from google.adk.streaming import StreamingMode

root_agent = LlmAgent(
    name="live_assistant",
    model="gemini-2.5-flash-live",   # Live variant
    instruction="Respond conversationally. Be concise; the user is speaking with you.",
    streaming_mode=StreamingMode.BIDI,
)

runner = Runner(agent=root_agent)

FastAPI WebSocket endpoint

from fastapi import FastAPI, WebSocket

app = FastAPI()

@app.websocket("/live")
async def live_endpoint(ws: WebSocket):
    await ws.accept()
    session = await runner.create_session(user_id="u1")

    async def producer():
        async for event in runner.run_live(session=session):
            await ws.send_json({"type": event.type, "data": event.serialize()})

    async def consumer():
        while True:
            msg = await ws.receive_json()
            await runner.send_live_input(session=session, input=msg)

    import asyncio
    await asyncio.gather(producer(), consumer())

Client (browser)

const ws = new WebSocket("ws://localhost:8000/live");
ws.onmessage = (e) => {
  const evt = JSON.parse(e.data);
  if (evt.type === "text_delta") appendText(evt.data.text);
  if (evt.type === "audio_chunk") playAudio(evt.data.bytes);
};
ws.send(JSON.stringify({ type: "user_text", text: "Hello!" }));

Response modalities

LlmAgent(
    ...,
    response_modalities=["TEXT", "AUDIO"],   # or just ["AUDIO"]
    voice_name="Aoede",                       # Gemini Live voice options
)

Validation

• WebSocket connection upgrades cleanly
• First text delta arrives < 500ms after user input
• Audio chunks decode without artifacts (test in browser)
• Session persists across reconnects (use a persistent SessionService)

See also

• audio-streaming-agent for voice-first patterns
• vision-streaming-agent for camera input
• bidirectional-tool-streaming for tools that run mid-stream