AI agent framework for Elixir built on OTP.

Agents follow The Elm Architecture (TEA): init/1, update/2, and optionally view/1. Each agent runs as a supervised OTP process with crash isolation, inter-agent messaging, and team coordination.

Quick Start

defmodule MyAgent do
  use Raxol.Agent

  def init(_ctx), do: %{findings: []}

  def update({:agent_message, _from, {:analyze, file}}, model) do
    {model, [shell("wc -l #{file}")]}
  end

  def update({:command_result, {:shell_result, %{output: out}}}, model) do
    {%{model | findings: [out | model.findings]}, []}
  end
end

{:ok, _} = Raxol.Agent.Session.start_link(app_module: MyAgent, id: :my_agent)
Raxol.Agent.Session.send_message(:my_agent, {:analyze, "lib/raxol.ex"})

Features

• TEA-based agents -- init/update/view with OTP supervision
• Agent teams -- Supervisor-based coordinator/worker groups
• Inter-agent messaging -- Registry-routed messages via Agent.Comm
• Real SSE streaming -- Anthropic, OpenAI, Ollama, Kimi, LLM7, OpenRouter, Lumo
• Shell commands -- Port-based execution with result callbacks
• Headless mode -- view/1 is optional; skip rendering entirely

Documentation

See the Agent Framework guide.