> ## Documentation Index
> Fetch the complete documentation index at: https://voxray-cac3ed72.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Anthropic

> Claude LLM provider for Voxray. Supports the full Claude 3 and Claude 4 model families as the LLM stage of the STT → LLM → TTS pipeline.

import { CardGroup, Card, Tabs, Tab, Warning, Note } from "@mintlify/components";

## Capabilities

<CardGroup cols={4}>
  <Card title="STT" icon="microphone" color="#6b7280">
    Not supported
  </Card>

  <Card title="LLM" icon="brain" color="#22c55e">
    Full Claude model family via the Messages API
  </Card>

  <Card title="TTS" icon="volume-high" color="#6b7280">
    Not supported
  </Card>

  <Card title="Realtime" icon="bolt" color="#6b7280">
    Not supported
  </Card>
</CardGroup>

Anthropic currently provides LLM capability only. Pair it with a supported STT provider (e.g. `"groq"` or `"openai"`) and TTS provider (e.g. `"elevenlabs"` or `"openai"`) to build a complete voice pipeline.

## API Key

Set `ANTHROPIC_API_KEY` as an environment variable or pass it inline under `api_keys` in `config.json`.

Get your key at [console.anthropic.com](https://console.anthropic.com).

## Quick Config

<Tabs>
  <Tab title="config.json">
    ```json theme={null}
    {
      "stt_provider": "groq",
      "llm_provider": "anthropic",
      "tts_provider": "elevenlabs",
      "model": "claude-haiku-4-5-20251001",
      "api_keys": {
        "anthropic": "sk-ant-...",
        "groq": "gsk_...",
        "elevenlabs": "..."
      }
    }
    ```
  </Tab>

  <Tab title="Environment variable">
    ```bash theme={null}
    export ANTHROPIC_API_KEY="sk-ant-..."
    export GROQ_API_KEY="gsk_..."
    export ELEVENLABS_API_KEY="..."
    ```

    Then in `config.json`:

    ```json theme={null}
    {
      "stt_provider": "groq",
      "llm_provider": "anthropic",
      "tts_provider": "elevenlabs",
      "model": "claude-haiku-4-5-20251001"
    }
    ```
  </Tab>
</Tabs>

## Available Models

| Model                       | Speed    | Cost    | Best For                                                                      |
| --------------------------- | -------- | ------- | ----------------------------------------------------------------------------- |
| `claude-haiku-4-5-20251001` | Fastest  | Lowest  | High-throughput voice agents, short-answer tasks, latency-sensitive workflows |
| `claude-sonnet-4-6`         | Balanced | Medium  | General-purpose voice agents, customer support, nuanced reasoning             |
| `claude-opus-4-6`           | Slowest  | Highest | Complex multi-step reasoning, analysis tasks where accuracy outweighs latency |

The default model when `model` is empty is `claude-3-sonnet-20240229` (defined as `DefaultLLMModel` in `pkg/services/anthropic/llm.go`).

## API Details

| Property       | Value                                                                        |
| -------------- | ---------------------------------------------------------------------------- |
| Endpoint       | `https://api.anthropic.com/v1/messages`                                      |
| Version header | `anthropic-version: 2023-06-01`                                              |
| Auth header    | `x-api-key: <your-key>`                                                      |
| Timeout        | 60 seconds per request                                                       |
| Max tokens     | 1024 per response (hardcoded default)                                        |
| Transport      | Shared `http.Transport` with 10 idle connections per host, 90 s idle timeout |

## Configuration Reference

| Key                  | Type   | Description                                                                                    |
| -------------------- | ------ | ---------------------------------------------------------------------------------------------- |
| `llm_provider`       | string | Set to `"anthropic"`                                                                           |
| `model`              | string | Claude model ID (e.g. `"claude-haiku-4-5-20251001"`); defaults to `"claude-3-sonnet-20240229"` |
| `api_keys.anthropic` | string | Anthropic API key (falls back to `ANTHROPIC_API_KEY`)                                          |
| `stt_provider`       | string | Required — Anthropic does not provide STT; use `"groq"`, `"openai"`, etc.                      |
| `tts_provider`       | string | Required — Anthropic does not provide TTS; use `"elevenlabs"`, `"openai"`, etc.                |

## Latency Consideration

<Warning>
  The Anthropic provider returns **full responses**, not streaming tokens. The implementation sets `"stream": false` in the API request and delivers the complete text as a single `LLMTextFrame` when the response arrives. This means TTS does not receive text incrementally — the first audio chunk has materially higher latency than streaming providers like OpenAI or Groq.

  To mitigate this:

  * Use shorter system prompts.
  * Prefer `claude-haiku-4-5-20251001` over larger models for faster time-to-first-token.
  * Consider sentence-level TTS batching if your pipeline supports it.
  * For latency-critical agents, evaluate `"llm_provider": "groq"` or `"llm_provider": "openai"`, both of which stream tokens as they are produced.
</Warning>

## Tool Calling

The Anthropic provider does **not** implement `LLMServiceWithTools`. MCP tool integration is not available with this provider. To use MCP tools, switch `llm_provider` to `"openai"` (the other provider that implements the tools interface).

## Notes and Limitations

* System messages in the conversation are concatenated and sent as the top-level `system` field in the Anthropic request; they are not included in the `messages` array.
* Messages with empty `content` are silently dropped before the request is sent.
* The `max_tokens` limit is hardcoded to 1024. Very long model outputs will be truncated. If your use case requires longer responses, this requires a code-level change in `pkg/services/anthropic/llm.go`.
* Only `"user"` and `"assistant"` roles are forwarded to Anthropic. Any message with a role other than `"system"` or `"assistant"` is treated as a user message.