Anthropic

This guide provides instructions for configuring and using Anthropic's Claude models within the AIGNE Framework via the @aigne/anthropic package. It covers API key setup, model instantiation, and invoking the model for both standard and streaming responses.

For a general overview of how models work within the AIGNE framework, please refer to the Models core concepts documentation.

Introduction#

The @aigne/anthropic package provides a direct and seamless integration between the AIGNE Framework and Anthropic's powerful Claude language models. This allows developers to leverage the advanced capabilities of models like Claude 3.5 Sonnet and Claude 3 Opus through the standardized ChatModel interface, ensuring consistency across your agentic applications.

Key features of this integration include:

• Direct API Integration: Utilizes the official Anthropic SDK for reliable communication.
• Chat Completions: Full support for Anthropic's chat completions API.
• Tool Calling: Natively supports Claude's tool-calling functionality.
• Streaming Responses: Enables real-time, responsive applications by handling streaming outputs.
• Type-Safe: Comes with comprehensive TypeScript typings for robust development.

Installation#

To get started, install the @aigne/anthropic package along with the core AIGNE package using your preferred package manager.

npm install @aigne/anthropic @aigne/core

yarn add @aigne/anthropic @aigne/core

pnpm add @aigne/anthropic @aigne/core

Configuration#

The AnthropicChatModel class is the primary entry point for interacting with Claude models. To instantiate it, you need to provide your Anthropic API key and optionally specify a model and other configurations.

API Key#

Your Anthropic API key can be configured in one of three ways, in order of precedence:

1. Directly in the constructor: Pass the key via the apiKey property.
2. ANTHROPIC_API_KEY environment variable: The model will automatically detect and use this variable.
3. CLAUDE_API_KEY environment variable: An alternative environment variable that is also supported.

import { AnthropicChatModel } from "@aigne/anthropic";

const model = new AnthropicChatModel({
  // Option 1: Provide the API key directly
  apiKey: "your-anthropic-api-key", 
  
  // The model will automatically use ANTHROPIC_API_KEY or CLAUDE_API_KEY
  // if they are set in your environment and apiKey is not provided.
});

Model Selection#

You can specify which Claude model to use with the model property. If not specified, it defaults to claude-3-7-sonnet-latest. Other common model parameters like temperature can be set within the modelOptions object.

A list of commonly used models includes:

• claude-3-5-sonnet-20240620
• claude-3-opus-20240229
• claude-3-sonnet-20240229
• claude-3-haiku-20240307

import { AnthropicChatModel } from "@aigne/anthropic";

const model = new AnthropicChatModel({
  apiKey: "your-anthropic-api-key",
  
  // Specify the model version
  model: "claude-3-haiku-20240307",

  // Configure other model behaviors
  modelOptions: {
    temperature: 0.7, // Controls randomness (0.0 to 1.0)
  },
});

Basic Usage#

To generate a response, use the invoke method. Pass a list of messages to the model to start a conversation. The method returns a promise that resolves with the model's output, including the text response and token usage statistics.

import { AnthropicChatModel } from "@aigne/anthropic";

const model = new AnthropicChatModel({
  apiKey: "your-anthropic-api-key",
  model: "claude-3-haiku-20240307",
});

const result = await model.invoke({
  messages: [{ role: "user", content: "Tell me about yourself" }],
});

console.log(result);

The result object contains the generated text and other metadata from the API.

Example Response

{
  "text": "I am Claude, a large language model trained by Anthropic.",
  "model": "claude-3-haiku-20240307",
  "usage": {
    "inputTokens": 8,
    "outputTokens": 12
  }
}

Streaming Responses#

For applications requiring real-time interaction, you can enable streaming by setting the streaming option to true in the invoke method. This returns an async iterator that yields response chunks as they become available.

The isAgentResponseDelta utility can be used to check if a chunk contains new data.

import { AnthropicChatModel } from "@aigne/anthropic";
import { isAgentResponseDelta } from "@aigne/core";

const model = new AnthropicChatModel({
  apiKey: "your-anthropic-api-key",
  model: "claude-3-haiku-20240307",
});

const stream = await model.invoke(
  {
    messages: [{ role: "user", content: "Tell me a short story about a robot." }],
  },
  { streaming: true },
);

let fullText = "";
const json = {};

for await (const chunk of stream) {
  if (isAgentResponseDelta(chunk)) {
    const text = chunk.delta.text?.text;
    if (text) {
      process.stdout.write(text); // Print text to the console as it arrives
      fullText += text;
    }

This code processes the stream, printing text chunks to the console immediately and accumulating the full response and metadata.

Summary#

You now have the necessary information to install, configure, and use Anthropic's Claude models within your AIGNE applications. You can perform basic invocations for simple tasks or use streaming for more interactive experiences.

To learn more about orchestrating multiple models and agents, see the Team Agent documentation. For details on other available models, visit the main Models section.