Hi! I’ve been building ClaudeCode, an SDK that lets you embed Claude as an agent in your Elixir apps - not just a chatbot, but an AI that can take actions.

The Problem

Most AI integrations are text-in/text-out. Building anything useful means parsing responses, manually wiring tool calls, handling conversation state, and mocking everything by hand for tests.

The Solution: Hermes + ClaudeCode

ClaudeCode wraps the claude cli, giving Claude full agentic capabilities. Combined with hermes_mcp v0.14.1 — Documentation , Claude can call your Elixir functions directly:

  # Define a tool
  defmodule MyApp.Tools.RefundOrder do
    use Hermes.Server.Component, type: :tool

    schema do
      field :order_id, :string, required: true
    end

    def execute(%{order_id: id}, frame) do
      {:ok, refund} = MyApp.Orders.refund(id)
      {:reply, Response.json(Response.tool(), refund), frame}
    end
  end

  # Create an MCP server
  defmodule MyApp.SupportTools do
    use Hermes.Server, name: "myapp", version: "1.0.0", capabilities: [:tools]

    component MyApp.Tools.RefundOrder
    component MyApp.Tools.OrderLookup
  end

  # Start a session
  {:ok, session} = ClaudeCode.start_link(mcp_servers: %{"myapp" => MyApp.SupportTools})

  # Ask it to do something
  session
  |> ClaudeCode.stream("Customer jane@example.com wants a refund for her last order")
  |> ClaudeCode.Stream.tap(&IO.inspect/1)
  |> ClaudeCode.Stream.final_text()

Claude will call OrderLookup, find Jane’s order, call RefundOrder, and explain what happened. No decision trees. No hardcoded workflows.

Testing Without API Calls

  test "support agent processes refunds" do
    ClaudeCode.Test.stub(ClaudeCode, fn _query, _opts ->
      [
        ClaudeCode.Test.tool_use("OrderLookup", %{email: "jane@example.com"}),
        ClaudeCode.Test.tool_result(%{id: "ORD-123", total: 99.99}]),
        ClaudeCode.Test.text("Processing refund for order ORD-123..."),
        ClaudeCode.Test.tool_use("RefundOrder", %{order_id: "ORD-123", reason: "request"}),
        ClaudeCode.Test.tool_result(%{status: "processed"}),
        ClaudeCode.Test.text("Done! Refunded $99.99.")
      ]
    end)

    {:ok, session} = ClaudeCode.start_link()
    result = session |> ClaudeCode.stream("Refund Jane") |> ClaudeCode.Stream.final_text()

    assert result =~ "Refunded $99.99"
  end

Millisecond tests. No API calls. Full async: true support.

Also Included

• OTP Supervision - Sessions restart on failure
• Native Streaming - Elixir Streams, LiveView-ready
• Multi-turn Memory - Context retained across queries
• Concurrent Agents - Run multiple sessions with the actor model

Links

• Hex: claude_code | Hex
• Docs: ClaudeCode v0.14.0 — Documentation
• GitHub:

Early days - I’d love to hear what agents you’d build, or what’s missing that would make this useful for your projects.