Custom server for live agents¶
The adk web tool runs a live agent for development purposes. It ships a browser client that captures the
microphone and camera, plays model audio, and renders transcripts, so you can talk to your
agent with no code of your own. Shipping to production means replacing that: running your own
server that bridges clients to run_live(), with the runner and session service initialized
once at startup and one LiveRequestQueue per connected user.
What follows is a complete FastAPI implementation of that bridge, and what a client needs to know to talk to it. It assumes you have read Sessions, which covers the lifecycle this example puts into practice.
FastAPI application example¶
This FastAPI application implements the bridge. It runs two concurrent tasks: an upstream
task that forwards WebSocket messages into LiveRequestQueue, and a downstream task that
forwards run_live() events back out.
import asyncio
from fastapi import FastAPI, WebSocket, WebSocketDisconnect
from google.adk.runners import Runner
from google.adk.agents.run_config import RunConfig
from google.adk.agents.live_request_queue import LiveRequestQueue
from google.adk.sessions import InMemorySessionService
from google.genai import types
from google_search_agent.agent import agent
# Application setup (once at startup)
APP_NAME = "live-agent"
app = FastAPI()
# Define your session service
session_service = InMemorySessionService()
# Define your runner
runner = Runner(
app_name=APP_NAME,
agent=agent,
session_service=session_service
)
@app.websocket("/ws/{user_id}/{session_id}")
async def websocket_endpoint(websocket: WebSocket, user_id: str, session_id: str) -> None:
await websocket.accept()
# Per-session setup: RunConfig, session, queue.
response_modalities = ["AUDIO"]
run_config = RunConfig(
response_modalities=response_modalities,
input_audio_transcription=types.AudioTranscriptionConfig(),
output_audio_transcription=types.AudioTranscriptionConfig(),
session_resumption=types.SessionResumptionConfig()
)
session = await session_service.get_session(
app_name=APP_NAME,
user_id=user_id,
session_id=session_id
)
if not session:
await session_service.create_session(
app_name=APP_NAME,
user_id=user_id,
session_id=session_id
)
live_request_queue = LiveRequestQueue()
async def upstream_task() -> None:
"""Receives messages from WebSocket and sends to LiveRequestQueue."""
try:
while True:
# Receive text message from WebSocket
data: str = await websocket.receive_text()
# Send to LiveRequestQueue
content = types.Content(parts=[types.Part(text=data)])
live_request_queue.send_content(content)
except WebSocketDisconnect:
# Client disconnected - signal queue to close
pass
async def downstream_task() -> None:
"""Receives Events from run_live() and sends to WebSocket."""
async for event in runner.run_live(
user_id=user_id,
session_id=session_id,
live_request_queue=live_request_queue,
run_config=run_config
):
# Send event as JSON to WebSocket
await websocket.send_text(
event.model_dump_json(exclude_none=True, by_alias=True)
)
# Run both tasks concurrently
try:
await asyncio.gather(
upstream_task(),
downstream_task(),
return_exceptions=True
)
finally:
live_request_queue.close() # Always close, even on error.
Async Context Required
All ADK bidirectional streaming applications must run in an async context. This requirement comes from multiple components:
run_live(): ADK's streaming method is an async generator with no synchronous wrapper (unlikerun())- Session operations:
get_session()andcreate_session()are async methods - WebSocket operations: FastAPI's
websocket.accept(),receive_text(), andsend_text()are all async - Concurrent tasks: The upstream/downstream pattern requires
asyncio.gather()for concurrent execution
All code examples assume an async context (within an async def or coroutine). They show the core logic without boilerplate wrapper functions.
Why two tasks¶
The bridge is two loops running at once, and that is what makes it bidirectional:
- Upstream reads from the WebSocket and pushes into the
LiveRequestQueue, so the user can send input at any moment, including while the agent is mid-sentence. - Downstream reads events from
run_live()and writes them to the WebSocket, streaming responses, transcriptions, and tool activity out as they happen.
Run them sequentially and you lose interruption: the server would be blocked reading the
agent's output while the user is trying to talk over it. asyncio.gather() is what keeps
both directions live simultaneously.
live_request_queue.close() must run on every exit path, including exceptions. An unclosed
queue leaves the Live API without a termination signal and can strand a session against your
concurrent-session quota until it times out, which is
what the try/finally is for.
gather(..., return_exceptions=True) collects exceptions rather than raising them, so check
the returned values if you need to distinguish a clean disconnect from a failure.
Production considerations¶
This example shows the core pattern. For production applications, consider:
- Error handling (ADK): Add proper error handling for ADK streaming events. For details on error event handling, see Error events.
- Handle task cancellation gracefully by catching
asyncio.CancelledErrorduring shutdown - Check exceptions from
asyncio.gather()withreturn_exceptions=True- exceptions don't propagate automatically
- Handle task cancellation gracefully by catching
- Error handling (Web): Handle web application-specific errors in upstream/downstream tasks. For example, with FastAPI you would need to:
- Catch
WebSocketDisconnect(client disconnected),ConnectionClosedError(connection lost), andRuntimeError(sending to closed connection) - Validate WebSocket connection state before sending with
websocket.client_stateto prevent errors when the connection is closed
- Catch
- Authentication and authorization: Implement authentication and authorization for your endpoints
- Rate limiting and quotas: Add rate limiting and timeout controls. For guidance on concurrent sessions and quota management, see Concurrent sessions.
- Structured logging: Use structured logging for debugging.
- Persistent session services: Consider using persistent session services (
DatabaseSessionServiceorVertexAiSessionService). See the ADK Session Services documentation for more details.
Connect a client¶
Your server exposes a WebSocket; something has to talk to it. During development that is
adk web. In production it is a client you write: a browser app, a mobile app, or a
telephony or WebRTC bridge. Whatever you build inherits the same contract, so it is worth
knowing exactly what adk web does and where it stops.
What adk web handles for you:
| Capability | What the built-in client does |
|---|---|
| Microphone | Captures and resamples to 16 kHz mono PCM, streamed as audio/pcm;rate=16000 |
| Playback | Plays model audio as 24 kHz mono PCM, gapless |
| Camera | Sends JPEG frames at ~1 fps as image/jpeg |
| Transcription | Renders both user and model transcripts, merging partial fragments |
| Barge-in | Stops playback when an event arrives with interrupted set |
What it does not do, and a production client may need:
- No screen sharing, and no video without an active audio call.
- No modality choice; responses are always
AUDIO. - No UI for proactivity, affective dialog, session resumption,
save_live_blob, or explicit VAD signals. Those are set on the server throughRunConfig. - No manual VAD; it relies on the server-side automatic detection that is on by default.
adk web and adk api_server both serve the same /run_live WebSocket; adk api_server
does not ship the browser client unless you pass --with_ui. You can therefore develop
against adk web and point a custom client at either.
Wire protocol and data format¶
The /run_live endpoint speaks JSON text frames only. Your client sends serialized
LiveRequest objects and receives serialized
Event objects. Binary data (audio and image bytes) is base64-encoded
inside the JSON, not sent as binary WebSocket frames.
On the client, branch on the same event fields you would in Python, in camelCase:
websocket.onmessage = (message) => {
const adkEvent = JSON.parse(message.data);
if (adkEvent.interrupted) {
stopAudioPlayback(); // user barged in; drop queued audio
finishCurrentBubble();
return;
}
if (adkEvent.turnComplete) {
finishCurrentBubble();
return;
}
for (const part of adkEvent.content?.parts ?? []) {
if (part.text) appendText(part.text);
if (part.inlineData) enqueueAudio(part.inlineData.data);
}
};
The media formats your client must produce and consume (sample rates, encodings, chunk
sizes) are in Audio and video. The streaming flags it branches on
(partial, turnComplete, interrupted) and how transcriptions fragment are in
Events.
Serializing events¶
The /run_live endpoint between ADK and the Live API is JSON-text-only, but the transport
between your server and your client is yours to design, and there you can send audio as
binary frames to avoid base64 overhead.
Event is a Pydantic model, so model_dump_json() converts it to a JSON string for a
WebSocket or SSE transport. Use by_alias=True for camelCase field names on the client and
exclude_none=True to drop empty fields:
async for event in runner.run_live(...):
await websocket.send_text(event.model_dump_json(exclude_none=True, by_alias=True))
Binary audio in inline_data is base64-encoded in JSON, which inflates the payload by about
33%. For audio-heavy streams, send audio as binary frames and metadata as JSON:
async for event in runner.run_live(...):
parts = event.content.parts if event.content else []
audio_parts = [p for p in parts if p.inline_data]
if audio_parts:
for part in audio_parts:
await websocket.send_bytes(part.inline_data.data)
# Metadata without the audio bytes.
await websocket.send_text(event.model_dump_json(
exclude={"content": {"parts": {"__all__": {"inline_data"}}}},
by_alias=True,
))
else:
await websocket.send_text(event.model_dump_json(exclude_none=True, by_alias=True))