[Feature Request] Support Using litellm.Router Instance

[Ocyss]

Is your feature request related to a problem? Please describe.

ADK currently supports using LiteLLM as a model provider. I noticed that while there is a llm_client parameter available, it does not seem to support passing a custom client instance. Moreover, the LiteLLM SDK features a Router that provides powerful load balancing capabilities, which could be particularly useful for managing multiple model deployments.

I prefer programmatic model management over environment variables, especially when working with multiple models that have different base URLs and API keys.

Describe the solution you'd like

I propose either:

1. Allowing ADK to globally accept a configured litellm.Router instance, or
2. Enabling litellm.Router to be passed directly via the llm_client parameter

Describe alternatives you've considered

1. LiteLLM Proxy - adds unnecessary external dependency

[kavinkumar807]

@Ocyss will check on the possiblities to implement this

[kavinkumar807]

@hangfei Looks like it is self assigned to you, is it still valid to contribute

[mdaffan01]

Is somebody working on this? It would be really helpful if this gets added.

[kavinkumar807]

@hangfei @boyangsvl if this is valid to contribute, shall I check with this ??

[hangfei]

@kavinkumar807 please help.

@selcukgun can provide some guidance.

[kavinkumar807]

@hangfei sure I'm happy to help.I will into this.

[anand-bakerev]

Hey team, is anyone working on this? This would be very helpful to build reliable agents. I have faced some issues in production because the model APIs were not available and hence impacting the google-adk agents as well. If router is supported then it can fallback to use different models when primary model APIs are not available (happened quite a few times with OpenAI and Gemini Both)