Proper way to integrate the ADK backend with custom frontend.

[unique1o1]

I'm following the ADK Streaming Quickstart guide, specifically Step 5 ("Building a Custom Streaming App"), to create a custom frontend interacting with an ADK agent using FastAPI and WebSockets.

The example here provided demonstrates setting up bidirectional streaming using runner.run_live which works well with the specified Gemini model (gemini-2.0-flash-live-001). The documentation also explicitly states that voice/video streaming (and by implication, the run_live function) requires Gemini models that support the Live API.

But my goal is to use a similar streaming setup (ADK backend, custom WebSocket frontend) but with a different Large Language Model (LLM), potentially with claude/openai or another custom model accessed via API, with no voice streaming.

When I try to configure the agent gpt-4o using litellm and use the runner.run_live function as shown in the custom app example, I encounter an error indicating that run_live is not implemented for that model type.

# From the example main.py - this call seems tied to Live API models
  File "/Users/yunikmaharjan/projects/adk_test/.venv/lib/python3.12/site-packages/google/adk/flows/llm_flows/base_llm_flow.py", line 82, in run_live
    async with llm.connect(llm_request) as llm_connection:
               ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/yunikmaharjan/projects/adk_test/.venv/lib/python3.12/site-packages/google/adk/models/base_llm.py", line 85, in connect
    raise NotImplementedError(
NotImplementedError: Live connection is not supported for openai/gpt-4o.

My question is: What is the recommended approach within the ADK framework to achieve real-time, bidirectional streaming between a custom frontend and the ADK backend.

Is there an alternative ADK function or pattern to handle streaming requests (LiveRequestQueue) and responses with different models?

[hangfei]

Right now, only the selected models are supported.

We are exploring how to support more models.

[unique1o1]

....potentially with claude/openai or another custom model accessed via API, with no voice streaming.

There is no support for other models without voice streaming, too?