Using Deepgram Voice Agent (Java SDK)

Run a full-duplex voice agent over a single WebSocket: user audio in, agent events + audio out.

When to use this product

• You want a live conversational agent.
• You need STT + think-provider + TTS orchestration in one session.
• You may need message injection, prompt updates, or function-call handling.

Use a different skill when:

• You only need transcription → deepgram-java-speech-to-text or deepgram-java-conversational-stt.
• You only need speech synthesis → deepgram-java-text-to-speech.
• You only need project/admin endpoints → deepgram-java-management-api.

Authentication

import com.deepgram.DeepgramClient;

DeepgramClient client = DeepgramClient.builder()
        .apiKey(System.getenv("DEEPGRAM_API_KEY"))
        .build();

The agent WebSocket uses the SDK's agent environment URL and the same auth headers.

Quick start

import com.deepgram.resources.agent.v1.types.AgentV1Settings;
import com.deepgram.resources.agent.v1.types.AgentV1SettingsAgent;
import com.deepgram.resources.agent.v1.types.AgentV1SettingsAgentThink;
import com.deepgram.resources.agent.v1.types.AgentV1SettingsAgentThinkOneItem;
import com.deepgram.resources.agent.v1.types.AgentV1SettingsAgentThinkOneItemProvider;
import com.deepgram.resources.agent.v1.types.AgentV1SettingsAudio;
import com.deepgram.resources.agent.v1.websocket.V1WebSocketClient;
import com.deepgram.types.OpenAiThinkProvider;
import java.util.List;
import java.util.Map;

V1WebSocketClient wsClient = client.agent().v1().v1WebSocket();

wsClient.onWelcome(welcome -> {
    OpenAiThinkProvider openAiProvider = OpenAiThinkProvider.of(Map.of("model", "gpt-4o-mini"));

    AgentV1Settings settings = AgentV1Settings.builder()
            .audio(AgentV1SettingsAudio.builder().build())
            .agent(AgentV1SettingsAgent.builder()
                    .think(AgentV1SettingsAgentThink.of(List.of(AgentV1SettingsAgentThinkOneItem.builder()
                            .provider(AgentV1SettingsAgentThinkOneItemProvider.of(openAiProvider))
                            .prompt("You are a helpful voice assistant. Keep your responses brief.")
                            .build())))
                    .greeting("Hello! How can I help you today?")
                    .build())
            .build();

    wsClient.sendSettings(settings);
});

wsClient.onConversationText(text -> System.out.printf("[%s] %s%n", text.getRole(), text.getContent()));
wsClient.onAgentStartedSpeaking(event -> System.out.println(">> Agent started speaking"));
wsClient.onAgentV1Audio(audioData -> System.out.printf("Received %d bytes%n", audioData.size()));

wsClient.connect().get(10, java.util.concurrent.TimeUnit.SECONDS);

Message injection / control

The repo also demonstrates:

wsClient.sendInjectUserMessage(com.deepgram.resources.agent.v1.types.AgentV1InjectUserMessage.builder()
        .content("What is the capital of France?")
        .build());

wsClient.sendInjectAgentMessage(com.deepgram.resources.agent.v1.types.AgentV1InjectAgentMessage.builder()
        .message("By the way, I can also help you with math and science questions!")
        .build());

Key parameters / API surface

• Connect path: client.agent().v1().v1WebSocket()
• Initial session config: AgentV1Settings
• Common send methods: sendSettings, sendMedia, sendUpdatePrompt, sendUpdateSpeak, sendInjectUserMessage, sendInjectAgentMessage, sendFunctionCallResponse, sendKeepAlive
• Event handlers: onWelcome, onSettingsApplied, onConversationText, onUserStartedSpeaking, onAgentThinking, onFunctionCallRequest, onAgentStartedSpeaking, onAgentAudioDone, onAgentV1Audio, onInjectionRefused, onPromptUpdated, onSpeakUpdated, onErrorMessage, onWarning
• Think-model discovery lives at client.agent().v1().settings().think().models().list()

API reference (layered)

1. In-repo source of truth: src/main/java/com/deepgram/resources/agent/v1/ and examples/agent/. No reference.md file is present.
2. Canonical AsyncAPI: https://developers.deepgram.com/asyncapi.yaml
3. Context7: /llmstxt/developers_deepgram_llms_txt
4. Product docs:
   • https://developers.deepgram.com/reference/voice-agent/voice-agent
   • https://developers.deepgram.com/docs/voice-agent
   • https://developers.deepgram.com/docs/configure-voice-agent
   • https://developers.deepgram.com/docs/voice-agent-message-flow

Gotchas

1. The base URL is the agent environment, not the standard API base. The SDK routes this automatically through environment().getAgentURL().
2. Send settings first. The repo examples wait for onWelcome(...) and immediately call sendSettings(...).
3. Audio is binary ByteString. Playback/output is your responsibility.
4. sendMedia(...) is raw audio bytes. Match whatever audio settings you configured.
5. Use the provider wrapper/union types rather than raw JSON. Constructors like OpenAiThinkProvider.of(...), AnthropicThinkProvider.of(...), GoogleThinkProvider.of(...) package the provider into the think/listen/speak union the SDK expects. The underlying payload is still an Object (so provider-field mistakes won't be caught at compile time), but the wrappers keep routing correct and ensure you pick the right variant of the sealed union.
6. There is no persisted agent-configuration management client shown in this checkout. This repo exposes live agent runtime plus think-model discovery.
7. Closing is connection-level. The examples call disconnect(); there is no separate close-message flow like Speak/Listen.

Example files in this repo

• examples/agent/VoiceAgent.java
• examples/agent/InjectMessage.java
• examples/agent/ProviderCombinations.java
• examples/agent/CustomProviders.java

Central product skills

For cross-language Deepgram product knowledge — the consolidated API reference, documentation finder, focused runnable recipes, third-party integration examples, and MCP setup — install the central skills:

npx skills add deepgram/skills

This SDK ships language-idiomatic code skills; deepgram/skills ships cross-language product knowledge (see api, docs, recipes, examples, starters, setup-mcp).