Extending Adapters

The true power of Agentic-Notify is its complete decoupling from external endpoints. The Orchestrator doesn't natively know how to speak to Google Calendar or Stripe—it relies entirely on Adapters.

Adapters are your physical tools. When you extend BaseAdapter, you automatically register that component with the internal ToolRegistry, making it instantly discoverable by the Agentic LLM Planner.

---

🛠️ Building a Custom Adapter

To build an interface with a new external system, simply subclass BaseAdapter and override the execute method.

Example: Creating a Stripe Payment Adapter

import logging
from typing import Dict, Any
from agentic_notify.adapters.base import BaseAdapter

logger = logging.getLogger(__name__)

class StripeRefundAdapter(BaseAdapter):
    """
    Adapter that interfaces with the Stripe API to issue refunds based on notification context.
    """

    @property
    def name(self) -> str:
        # This is the exact keyword the LLM and the Workflow schema will call
        return "issue_stripe_refund"

    async def execute(self, payload: Dict[str, Any]) -> Dict[str, Any]:
        charge_id = payload.get("charge_id")
        amount = payload.get("amount")

        if not charge_id:
            raise ValueError("StripeRefundAdapter requires a 'charge_id'.")

        logger.info(f"Issuing refund for {charge_id} handling {amount} cents.")

        # Insert real `stripe.Refund.create(charge=charge_id)` logic here

        return {
            "status": "success",
            "refund_id": "re_12345",
            "processed_amount": amount
        }

🔌 Registering Your Adapter

Once defined, you must mount it to the active NotificationOrchestrator before starting your integrations (like FastAPI).

from agentic_notify.orchestrator import NotificationOrchestrator
from my_custom_code import StripeRefundAdapter

orchestrator = NotificationOrchestrator(...)

# Register it dynamically
orchestrator.register_adapter(StripeRefundAdapter())

Once registered, if your AgenticPlanner encounters an intent like "A customer complained about charge ch_999, issue a refund", the LLM will automatically generate a step mapped to issue_stripe_refund and map the charge ID into the payload!