Backend for native Elixir module execution.

Executes skills implemented as Elixir modules directly in the BEAM, providing type-safe, in-process execution with full access to the application's runtime context.

Usage

defmodule MyApp.Skills.Database do
  @behaviour Conjure.NativeSkill

  def __skill_info__ do
    %{
      name: "database",
      description: "Query the database",
      allowed_tools: [:execute, :read]
    }
  end

  def execute(query, _context), do: {:ok, run_query(query)}
  def read(table, _context, _opts), do: {:ok, get_schema(table)}
end

session = Conjure.Backend.Native.new_session([MyApp.Skills.Database], [])

{:ok, response, session} = Conjure.Backend.Native.chat(
  session,
  "What tables do we have?",
  &api_callback/1,
  []
)

How It Works

1. Native skill modules implement Conjure.NativeSkill behaviour
2. This backend generates Claude tool definitions from skill info
3. When Claude returns tool_use blocks, this backend:
   • Maps tool names back to skill modules
   • Invokes the appropriate callback (execute/read/write/modify)
   • Returns results to Claude

Advantages Over Local Backend

• No subprocess/shell overhead
• Type-safe with compile-time checks
• Direct access to application state (Ecto repos, caches, GenServers)
• Better error handling with pattern matching

Options

• :working_directory - Working directory for file operations
• :timeout - Execution timeout in milliseconds (default: 30_000)
• :max_iterations - Maximum tool-use iterations (default: 25)

See Also

• Conjure.NativeSkill - Behaviour for skill modules
• Conjure.Backend.Local - Shell-based execution