Anthropic (Claude)

Use AnthropicLLM when you want Claude’s strong instruction-following and reasoning for your voice assistant.

Prerequisites

• An Anthropic API key or a CompositeVoice proxy server
• No additional dependencies required. AnthropicLLM uses native fetch internally.

Basic setup

import { CompositeVoice, AnthropicLLM, NativeSTT, NativeTTS } from '@lukeocodes/composite-voice';

const agent = new CompositeVoice({
  providers: [
    new NativeSTT({ language: 'en-US' }),
    new AnthropicLLM({
      proxyUrl: '/api/proxy/anthropic',
      model: 'claude-haiku-4-5',
      maxTokens: 512,
      systemPrompt: 'You are a concise voice assistant. Keep answers under two sentences.',
    }),
    new NativeTTS(),
  ],
});

await agent.initialize();
await agent.startListening();

Configuration options

Model variants

Complete example

import {
  CompositeVoice,
  MicrophoneInput,
  AnthropicLLM,
  DeepgramSTT,
  DeepgramTTS,
  BrowserAudioOutput,
} from '@lukeocodes/composite-voice';

const agent = new CompositeVoice({
  providers: [
    new MicrophoneInput(),
    new DeepgramSTT({
      proxyUrl: '/api/proxy/deepgram',
      language: 'en',
      options: { model: 'nova-3', smartFormat: true },
    }),
    new AnthropicLLM({
      proxyUrl: '/api/proxy/anthropic',
      model: 'claude-haiku-4-5',
      maxTokens: 256,
      temperature: 0.7,
      systemPrompt: 'You are a friendly voice assistant. Answer briefly.',
    }),
    new DeepgramTTS({
      proxyUrl: '/api/proxy/deepgram',
      voice: 'aura-2-thalia-en',
    }),
    new BrowserAudioOutput(),
  ],
  conversationHistory: { enabled: true, maxTurns: 10 },
});

await agent.initialize();
await agent.startListening();

Tool use / function calling

AnthropicLLM implements the ToolAwareLLMProvider interface, enabling the LLM to call functions during generation. Text output streams to TTS as usual, while tool calls are handled via an async callback. After tool execution, the SDK automatically re-calls the LLM with the tool result so it can generate a natural language follow-up.

Configure tools on the top-level CompositeVoice config:

const voice = new CompositeVoice({
  providers: [
    new MicrophoneInput(),
    new DeepgramSTT({ proxyUrl: '/api/proxy/deepgram' }),
    new AnthropicLLM({ proxyUrl: '/api/proxy/anthropic', model: 'claude-sonnet-4-6' }),
    new DeepgramTTS({ proxyUrl: '/api/proxy/deepgram' }),
    new BrowserAudioOutput(),
  ],
  tools: {
    definitions: [
      {
        name: 'get_weather',
        description: 'Get the current weather for a location',
        parameters: {
          type: 'object',
          properties: {
            location: { type: 'string', description: 'City name' },
          },
          required: ['location'],
        },
      },
    ],
    onToolCall: async (toolCall) => {
      if (toolCall.name === 'get_weather') {
        const weather = await fetchWeather(toolCall.arguments.location);
        return { toolCallId: toolCall.id, content: JSON.stringify(weather) };
      }
      return { toolCallId: toolCall.id, content: 'Unknown tool' };
    },
  },
});

Tool calls are transparent to the event stream — you still receive the same llm.chunk and llm.complete events. The tool execution loop runs internally: when the LLM emits a tool_use stop reason, the SDK calls your onToolCall handler, appends the result to the conversation, and re-invokes the LLM.

Note: Tool use is supported by AnthropicLLM and all OpenAI-compatible providers (OpenAILLM, GroqLLM, GeminiLLM, MistralLLM, OpenAICompatibleLLM). WebLLMLLM does not support tools.

Tips

• maxTokens is required. Anthropic’s API rejects requests without it. The provider defaults to 1024, but for voice applications, lower values (128—512) produce faster responses.
• System prompts are handled automatically. AnthropicLLM extracts system messages and passes them as the top-level system parameter, which is what the Anthropic API expects.
• Use a proxy in browsers. The proxy server injects your API key server-side so it never reaches the client.
• Streaming is on by default. Set stream: false only if you need the complete response before TTS begins.

Related

• Providers reference — all LLM providers at a glance
• API reference — full class documentation
• Configuration guide — global CompositeVoice settings