Gemini Live API Support

[imrenagi]

Is your feature request related to a specific problem?

While Gemini API Live (bidirectional streaming) capabilities is available and documented for the Python ADK (as seen here: https://google.github.io/adk-docs/streaming/#supported-models), I'm not sure whether we can utilize these real-time, low-latency voice and multimodal interaction features natively within the Golang ADK ecosystem. Do we have this capability right now? Any pointer how to use it in go?

Impact on your work

This feature will be greate to have in golang adk for building real-time AI agents and voice interfaces.

---

🟡 Recommended Information

Willingness to contribute

• Yes, I would be interested in contributing a PR (with guidance).
• No, I cannot contribute at this time.

[kbpranay]

Native Gemini Live support is great for latency, but the trade-off in the current raw API is the lack of granular controls for things like barge-in and custom VAD logic.

We've been working on this with eboo.ai, where we use a custom pipeline to manage the orchestration rather than just wrapping the SDK. This helps minimize the latency spikes Gemini Live can hit and gives back control over how interruptions are handled—which is critical for making it feel like a real conversation. If you're running into issues with the default turn-taking logic, I'd be happy to discuss the pipeline approach we're using.

[imrenagi]

Hi @kbpranay , yes it would be great if you can share the approach. I was thinking earlier that default turn-taking logic in the existing gemini live will just be enough for live api mvp in adk. so i took a look into how it was implemented in python and some dev guideline in adk docs. If there is addition or improvement to that that i can consider, that would be great.

just want to clarify, the custom pipeline you are currently working on, is it going to be the addition for adk-go quite soon? Thanks!

[kbpranay]

While native Go support for the Gemini Live API is in the works, we've found that simply 'wrapping' the WebSocket isn't enough for production voice agents. At eboo.ai, we developed a custom orchestration pipeline to solve the granular control issues often hit with the raw SDKs—specifically around VAD/barge-in logic and telephony bridging. We found that the standard 'barge-in' often feels jarring; our pipeline uses a custom VAD logic to handle interruptions much more naturally. If anyone here is struggling with these orchestration 'dirty work' details while waiting for the official Go ADK implementation, we've got some patterns that might help.

[kbpranay]

If you are looking to bridge Gemini Live to telephony or need more orchestration control, check out how eboo.ai handles this. We actually built a custom pipeline that solves the standard SDK limitations—specifically around barge-in logic and reducing the orchestration gap when bridging to WebRTC/telephony. It moves the complex orchestration out of your Go app and into an optimized voice pipeline.