ADK-Rust

Realtime Examples

Four runnable examples in the examples/ directory, in rough order of how much they teach. All use the server-side bridge; all run on either OpenAI or Gemini. Set the relevant key first:

export OPENAI_API_KEY=sk-…        # for OpenAI sessions
export GEMINI_API_KEY=…           # for Gemini sessions (or GOOGLE_API_KEY)

realtime_tools

Headless. The clearest first read. A console probe (no browser) that exercises tool dispatch on IntegratedRealtimeRunner: single-tool, parallel-tool, and calculator turns, with session/memory/transcript aggregation wired in. The smallest end-to-end picture of how a tool turn flows.

cargo run --manifest-path examples/realtime_tools/Cargo.toml

Start here to understand the runner; then move to the web examples for audio.

realtime_voice — Mindfulness with Mia

The flagship. A full web voice app: browser captures mic PCM16 and plays the agent gaplessly with barge-in; the Rust server owns the session. Its standout feature is memory: a file-backed GraphMemoryService is Mia's long-term memory — her profile card is injected at session start, every turn is logged to the graph, and she curates durable facts mid-conversation via the bridged remember/relate tools. A live "User Memory Insights" panel reads and writes the same graph over /api/memory. Also shows a server-side get_weather tool.

LightDark
Mindfulness with Mia — light themeMindfulness with Mia — dark theme
cargo run --manifest-path examples/realtime_voice/Cargo.toml
# → http://localhost:3033

Read it for: Web Audio capture/playback, knowledge-graph memory, agent self-curation.

customer_service — Aria

The multimodal showcase. A redesigned take on Google's "Customer Service Agent" Live API demo — backend-agnostic, in Rust, with a themed (system/light/dark) three-column UI. It combines everything:

  • Multimodal — streams mic PCM and camera JPEG frames; Aria sees what you show her (send_video_frame).
  • Tools — process_refund and connect_to_human run server-side; results are spoken back.
  • Affective dialogue — empathetic on both backends; CS_AFFECTIVE=1 switches Gemini to the native-audio model for true tone-matching.
LightDark
Customer Service Agent — light themeCustomer Service Agent — dark theme
cargo run --manifest-path examples/customer_service/Cargo.toml
# → http://localhost:3066

# headless smoke test (verifies the refund tool runs):
cargo run --manifest-path examples/customer_service/Cargo.toml -- probe openai
cargo run --manifest-path examples/customer_service/Cargo.toml -- probe gemini

# enable Gemini native affective dialogue:
CS_AFFECTIVE=1 cargo run --manifest-path examples/customer_service/Cargo.toml

Read it for: video input, mixing multimodal + tools + affect in one agent.

live_translation

The dedicated-model example. A real-time speech-to-speech translator: speak English, hear Spanish a couple of seconds later. Same bridge, but pointed at the providers' translation models — OpenAI gpt-realtime-translate and Gemini gemini-3.5-live-translate-preview — which speak a different protocol than the conversational models. Pick a target language in the UI; the server negotiates per-provider audio rates before audio flows.

LightDark
Live Translation — light themeLive Translation — dark theme
cargo run --manifest-path examples/live_translation/Cargo.toml
# → http://localhost:3055
cargo run --manifest-path examples/live_translation/Cargo.toml -- probe openai

Read it for: using the dedicated translation models and a focused, minimal bridge.

A suggested path

  1. realtime_tools — see the runner and a tool turn, no audio noise.
  2. realtime_voice — add real audio and knowledge-graph memory.
  3. customer_service — add vision and affect; the full multimodal agent.
  4. live_translation — a specialized model and protocol.

Each README.md in the example directory has the full env-var table, model overrides (OPENAI_REALTIME_MODEL / GEMINI_REALTIME_MODEL), and architecture notes.

← Back to the Realtime overview