Anthropic Elixir API Wrapper

This unofficial Elixir wrapper provides a convenient way to interact with the Anthropic API, specifically designed to work with the Claude LLM model. It includes modules for handling configuration, preparing requests, sending them to the API, and processing responses.

Features

• Easy setup and configuration.
• Support for registering and invoking tools.
• Support for sending messages and receiving responses from the Claude LLM model.
• Error handling for both client and server-side issues.
• Customizable request parameters to tweak the behavior of the API.

To-dos

• Add streaming handling
• Allow for Tool do define different models.

Installation

The package can be installed by adding anthropic to your list of dependencies in mix.exs:

def deps do
  [
    {:anthropic, "~> 0.4.0", hex: :anthropic_community}
  ]
end

Configuration

Add or create a config file to provide the api_key as in the example below. Or pass an :api_key as option on Anthropic.new(api_key: "key")

# config/config.exs
import Config

config :anthropic,
  api_key: System.get_env("ANTHROPIC_API_KEY")

Usage

Basic Conversation

{:ok, response, request} =
  Anthropic.new()
  |> Anthropic.add_system_message("You are a helpful assistant")
  |> Anthropic.add_user_message("Explain monads in computer science. Be concise.")
  |> Anthropic.request_next_message()

The response will hold a map with the API response.

%{
  id: "msg_013Zva2CMHLNnXjNJJKqJ2EF",
  content: [
    %{type: "text", content: "Monads in computer science are a concept borrowed from category theory in mathematics, applied to abstract and manage complexity in functional programming. They provide a framework for chaining operations together step by step, where each step is processed in a context that can handle aspects like computations with side effects (e.g., state changes, I/O operations), errors, or asynchronous operations, without losing the purity of functional programming."}
  ],
  model: "claude-3-opus-20240229",
  role: "assistant",
  stop_reason: "end_turn",
  stop_sequence: null,
  type: "message",
  usage: %{
    "input_tokens" => 10,
    "output_tokens" => 25
  }
}

The conversation can continue:

request
|> Anthropic.add_user_message("Hold on right there! ELI5!")
|> Anthropic.request_next_message()

Registering and Invoking Tools

You can register tools that the AI can use to perform specific tasks. Here's an example:

defmodule MyApp.WeatherTool do
  @behaviour Anthropic.Tool.ToolBehaviour


  @impl true
  def description do
    """
    Tool description to explain AI how and when to use it
    """
  end

  @impl true
  def parameters do
    [
      {:location, :string, "The city and state of the location you need the weather"}
    ]
  end

  @impl true
  def invoke(location: location) do
    # Implement the tool's functionality here
    # ...
    # and return a string
  end
end

{:ok, response, request} =
  Anthropic.new()
  |> Anthropic.register_tool(MyApp.WeatherTool)
  |> Anthropic.add_user_message("Use the MyApp.WeatherTool to perform a task.")
  |> Anthropic.request_next_message()
  |> Anthropic.process_invocations()