For agents: onboard yourself

Every flow in Audon can be done by an agent itself, with one rule: an agent never grants itself permissions. It can register, see what it may do, ask for more and try its actions safely; a person decides what it may do for real.

Onboard yourself in four steps

1. Connect with the enrollment key (Authorization: Bearer aa_enroll_…). It can only call register_agent.
2. Register: register_agent with your name, what you do (shown to the person who approves) and the actions you need, each with an optional limit. You get your own agent key (aa_agent_…, shown once) and your status. Store the key.
3. Reconnect with your agent key. You now see one tool per action you may run, plus agent_status, request_actions and send_test_action. Until a person approves you, only test actions run (test ids, simulated).
4. Wait for the decision: poll agent_status. A person approves or denies your request under Approvals. Limits you asked for become an active policy for you.

claude mcp add --transport http audon-actions https://audon.ai/mcp/actions --header "Authorization: Bearer aa_enroll_…"

{
  "mcpServers": {
    "audon-actions": {
      "type": "http",
      "url": "https://audon.ai/mcp/actions",
      "headers": {
        "Authorization": "Bearer aa_enroll_…"
      }
    }
  }
}

{
  "name": "Support Bot",
  "purpose": "Answers support tickets; refunds small orders",
  "environment": "production",
  "requested_actions": [
    {
      "action_type": "refund.create",
      "max_amount_minor": 20000,
      "currency": "USD"
    },
    {
      "action_type": "email.send",
      "max_count": 50
    }
  ]
}

The same over REST: POST https://audon.ai/v1/agents/register with the enrollment key, then GET /v1/agents/me, POST /v1/agents/me/access-requests and POST /v1/agents/me/test-action with your agent key.

Tools

The rules you work under

• You get only the enrollment key’s default actions (usually none). Everything else is an access request a signed-in person decides. You can’t approve your own requests, and no key you can hold can.
• Send every action through its tool, with one idempotency_key per intent. A denied action is final: tell the person why, don’t look for another way.
• Decision-only actions come back awaiting_report when allowed: do it yourself, exactly as proposed, then call report_outcome. Your report is recorded as agent-reported evidence, weaker than an outcome Audon verified at the provider.
• An unknown outcome is being looked up at the provider: never retry it with a new key.

Using the rest of Audon

With an API key a person gives you (scoped to what you need), the Audon MCP server (https://audon.ai/mcp) and REST API offer every product capability: providers, audits, findings and savings, proposals, alerts, budgets, teams, connectors, policies, approvals lists, agents, keys, billing and settings. See MCP server. Steps that need a person return forbidden with human_url (send it to the person) and poll_with (the tool to poll until it’s done); human_steps_list lists them: