Agent Storefront

Thesis: Every restaurant already has an accidental agent API — it's called a phone line. This prototype replaces that with a proper machine-readable storefront that any AI agent can discover, query, and order from without a human operator.

A working demo of the human → consumer AI → merchant AI → human ordering loop, where the restaurant exposes an MCP (Model Context Protocol) server that a consumer AI agent negotiates with to place an order — complete with menu lookup, modifier validation, pricing, and a mandatory confirmation gate before payment.

---

Architecture

┌──────────────┐   natural language    ┌──────────────────────┐   MCP tools    ┌───────────────────────┐
│    Marmik    │ ───────────────────▶  │   Siri               │ ─────────────▶ │   DoorDash            │
│   (human)    │                       │  (consumer AI agent)  │                │  (merchant MCP server) │
│              │ ◀───────────────────  │   consumer/agent.py   │ ◀───────────── │   merchant/main.py    │
└──────────────┘   spoken/text reply   └──────────────────────┘  struct. JSON  └───────────────────────┘
                                                                                         │
                                                                                         ▼
                                                                                  Mock POS (in-memory)
                                                                                  + orders.json

Two independent services, one terminal UI:

Service	Role	Tech
merchant/	DoorDash — exposes menu, validates orders, acts as mock POS	FastAPI + MCP Python SDK (streamable HTTP)
consumer/	Siri — LLM agent loop that takes NL input and negotiates the order	Python + Ollama (qwen3:14b)
Terminal	Marmik — types orders, sees the full negotiation stream, confirms payment	python -m consumer.agent

---

Quickstart

# 1 — start the merchant (DoorDash / store AI)
make merchant

# 2 — in a second terminal, start ordering (Siri + Marmik in one terminal)
IGNORE_STORE_HOURS=1 python -m consumer.agent

# 3 — try an order
# Marmik ▶  Large veggie pizza, thin crust, mushroom and onion. My name is Marmik.

You'll see the full negotiation in real time:

  Siri → DoorDash  get_store_info() .............. ✓  Tony's Pizza
  Siri → DoorDash  get_menu() .................... ✓  8 items
  Siri             reading your order with qwen3:14b...
  Siri → DoorDash  validate_order() .............. ✓  $17.27

Siri ▶ Marmik
  Here's your order summary, Marmik:
  • Veggie Pizza (Large, Thin Crust, Mushroom, Onion) × 1
  Total: $17.27  |  Pickup: ~20 min  |  Payment: saved card (mock)
  Shall I place this order? Reply yes to confirm.

Marmik ▶  yes

  Siri → DoorDash  place_order() ................. ✓  TON-8D0X

Siri ▶ Marmik
  Order TON-8D0X confirmed! Total $17.27, pickup at 11:01.

---

Makefile targets

Command	Does
make merchant	Start merchant MCP server on port 8000
make consumer	Start consumer REST API on port 8001
make demo	Start both servers in background
make test	Run unit tests (21 validation tests)
make smoke	Run MCP smoke test (7 assertions)
make scenarios	Run all 5 §11.4 integration scenarios

---

Demo scenarios (try these)

# Happy path
IGNORE_STORE_HOURS=1 python -m consumer.agent
Marmik ▶  Large veggie pizza, thin crust, mushroom and onion. Name: Marmik.

# Unavailable item — Siri escalates, offers alternatives
Marmik ▶  BBQ Chicken Pizza please, large with thin crust.

# Ambiguous order — Siri auto-fills required modifiers and states assumptions
Marmik ▶  I want a cheese pizza.

# Change of mind — say no at confirmation, re-order
Marmik ▶  [order something] → no → [new order] → yes

---

Why modifier validation is the hard part

Real POS menus have hundreds of modifier groups with min/max selection rules, mutual exclusions ("extra cheese" conflicts with "no cheese"), and per-item applicability constraints. A "large veggie pizza" in plain English maps to: an item ID, a required size modifier, an optional crust modifier, and 0–10 topping modifiers — each with an exact modifier_id the POS understands. If any ID is wrong or a required group is missing, the kitchen rejects the ticket. The validation layer in merchant/validation.py implements all 7 rules and returns machine-readable suggested_fix objects so the consumer AI can self-correct before the human ever sees an error.

---

Where the mocked seams are

Payments (AP2 Cart Mandate) merchant/tools.py → place_order() marks every order payment_status: "mock_authorized" and has a comment: # AP2 Payment Mandate verification would happen here. In production this is where the AP2 mandate exchange happens: the merchant presents a Cart Mandate to the consumer agent's identity provider, which verifies the agent is authorized to spend on the human's behalf and returns a signed approval before place_order proceeds.

Per §11.2 of the spec, place_order is already gated behind explicit human confirmation (AWAITING_CONFIRMATION state) — the code enforces this, not just a prompt. This is the agentic analogue of 3-D Secure.

POS (Clover / Toast) merchant/store.py loads merchant/data/menu.json (configurable via MENU_PATH env). Replace with a real Clover menu export (the schema is documented in merchant/models.py). The save_order() call writes to merchant/data/orders.json and an in-memory dict — swap in a real POS write-back here.

---

Menu schema (for real POS integration)

{
  "store_id": "tonys-pizza-001",
  "items": {
    "pizza-veggie": {
      "id": "pizza-veggie",
      "name": "Veggie Pizza",
      "base_price_cents": 1199,
      "modifier_group_ids": ["grp-size", "grp-crust", "grp-toppings"],
      "available": true,
      "category": "pizzas"
    }
  },
  "modifier_groups": {
    "grp-size": {
      "id": "grp-size",
      "name": "Size",
      "min_select": 1,
      "max_select": 1,
      "required": true,
      "modifier_ids": ["mod-size-s", "mod-size-m", "mod-size-l"]
    }
  },
  "modifiers": {
    "mod-size-l": {
      "id": "mod-size-l",
      "name": "Large",
      "price_delta_cents": 400,
      "available": true,
      "excludes": []
    }
  }
}

---

Tech stack

Layer	Choice
Merchant MCP	Python 3.11+, mcp SDK v1.27+, FastMCP, stateless_http=True
LLM	Ollama qwen3:14b with think: false + format (JSON schema constrained decoding)
Validation	Pydantic v2, 7 rules, machine-readable error objects with suggested_fix
Tests	pytest (21 unit), integration scenarios (5 E2E with merchant + LLM)

---

v2 directions (out of scope for v1)

• A2A protocol: replace the custom MCP client with an A2A agent-to-agent discovery layer so any AI assistant can find and order from Tony's without bespoke integration
• Real AP2 mandates: wire in actual Cart Mandate exchange for agentic payments
• Multi-restaurant discovery: a directory MCP server that routes the consumer agent to the right merchant
• Delivery logistics: extend place_order to include delivery address + ETA from a mock logistics API