deepgram-js-voice-agent

Using Deepgram Voice Agent (JavaScript / TypeScript SDK)

Full-duplex voice agent runtime over wss://agent.deepgram.com/v1/agent/converse: audio in, LLM orchestration, audio out, plus function calling and prompt/runtime updates.

When to use this product

• You want an interactive voice assistant where the user speaks, the agent thinks, and the agent responds with speech.
• You need function / tool calling inside the conversation loop.
• You want Deepgram to host the STT + think + TTS orchestration.

Use a different skill when:

• You only need transcription → deepgram-js-speech-to-text or deepgram-js-conversational-stt.
• You only need synthesis → deepgram-js-text-to-speech.
• You want project keys, usage, models, or other admin APIs → deepgram-js-management-api.

Authentication

require("dotenv").config();

const { DeepgramClient, DeepgramEnvironment } = require("@deepgram/sdk");

const deepgramClient = new DeepgramClient({
  apiKey: process.env.DEEPGRAM_API_KEY,
  environment: DeepgramEnvironment.Agent,
});

The websocket itself is routed to the agent host by src/CustomClient.ts, but the repo example uses DeepgramEnvironment.Agent so client.agent.v1.settings.think.models.list() also points at the agent base.

Quick start

From examples/09-voice-agent.ts:

const deepgramConnection = await deepgramClient.agent.v1.createConnection();

deepgramConnection.on("message", (data) => {
  if (data.type === "ConversationText") {
    console.log("Conversation text:", data);
  } else if (typeof data === "string") {
    console.log("Audio received (length):", data.length);
  }
});

deepgramConnection.connect();
await deepgramConnection.waitForOpen();

deepgramConnection.sendSettings({
  type: "Settings",
  audio: {
    input: { encoding: "linear16", sample_rate: 24000 },
    output: { encoding: "linear16", sample_rate: 16000, container: "wav" },
  },
  agent: {
    language: "en",
    listen: { provider: { type: "deepgram", model: "nova-3" } },
    think: {
      provider: { type: "open_ai", model: "gpt-4o-mini" },
      prompt: "You are a friendly AI assistant.",
    },
    speak: { provider: { type: "deepgram", model: "aura-2-thalia-en" } },
    greeting: "Hello! How can I help you today?",
  },
});

The same example also shows client.agent.v1.settings.think.models.list() for discovering supported think models.

Key parameters / API surface

• Connection setup: client.agent.v1.createConnection() / connect().
• First outbound control message: sendSettings(AgentV1Settings).
• Runtime updates: sendUpdatePrompt(...), sendUpdateThink(...), sendUpdateSpeak(...), sendInjectUserMessage(...), sendInjectAgentMessage(...), sendFunctionCallResponse(...), sendKeepAlive(...), sendMedia(...).
• Important inbound events from src/api/resources/agent/resources/v1/client/Socket.ts: Welcome, SettingsApplied, ConversationText, UserStartedSpeaking, AgentThinking, FunctionCallRequest, AgentStartedSpeaking, AgentAudioDone, Warning, Error, plus audio payloads.

Limitations

This SDK exposes the live agent runtime plus settings.think.models.list(), but it does not expose persisted Voice Agent configuration CRUD endpoints in the current generated surface.

API reference (layered)

1. In-repo reference: reference.md → Agent V1 Settings Think Models; live websocket behavior is defined in src/CustomClient.ts and src/api/resources/agent/resources/v1/client/{Client,Socket}.ts.
2. Canonical OpenAPI (REST): https://developers.deepgram.com/openapi.yaml
3. Canonical AsyncAPI (WSS): https://developers.deepgram.com/asyncapi.yaml
4. Context7: library ID /llmstxt/developers_deepgram_llms_txt
5. 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. Settings must be first. Send sendSettings({ type: "Settings", ... }) immediately after the socket opens.
2. Audio and JSON events share the same message stream. Your message handler must branch on typeof data and data.type.
3. Keepalive matters. examples/09-voice-agent.ts sends KeepAlive every 5 seconds to preserve long sessions.
4. Encoding/sample rates must line up on both sides. Mismatches cause distorted uploads or unusable playback.
5. Think-model discovery is separate from the websocket. Use client.agent.v1.settings.think.models.list() before choosing providers.
6. Function-call requests arrive as arrays. Inspect data.functions[], then answer with sendFunctionCallResponse({ type: "FunctionCallResponse", id, name, content }).
7. Persisted agent configurations are not in this SDK today. If you need stored configs, use raw HTTP or another SDK surface.

Example files in this repo

• examples/09-voice-agent.ts
• examples/34-agent-custom-providers.ts
• examples/35-agent-provider-combinations.ts
• examples/36-agent-inject-message.ts

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).