Supplier coordination lives in chat threads
Restock timing, parcel counts and domestic tracking numbers move through WeChat and email. An agent cannot reliably parse a workflow that exists as screenshots.
Guide · AI-run stores
Your agent can run the store. Can it run fulfillment?
AI agents now build Shopify and WooCommerce stores, write the listings, set the prices and answer the customers. Then somebody orders a physical product from a supplier in China, and the agent hits the one part of e-commerce that still runs on chat threads, spreadsheets and PDFs. This guide covers what breaks, what an agent-operable fulfillment layer has to provide, and a working implementation over MCP.
Last updated: 2026-07-27
Where AI-run stores hit the wall
Quotes arrive as spreadsheets
Freight pricing is quoted ad hoc, per conversation. Without a structured estimate call, an agent cannot compare options or detect drift between quote and bill.
Tracking is scattered across carriers
The China domestic leg, warehouse handling and the international carrier each speak a different system. The agent ends up polling six tabs a human used to keep open.
Billing is a PDF at month-end
If charges cannot be reconciled line-by-line against a quote by a program, the agent must either trust blindly or escalate everything — which is not automation.
What “agent-operable fulfillment” requires
These are protocol-level properties, not dashboard features. If any of them is missing, the agent degrades into a screen-scraper with an API key:
- Typed tool calls for every operational step — estimate, order, instruct, track, settle — not web pages a human must read
- Idempotent writes: a retried order-creation returns the original response instead of creating a duplicate shipment
- HMAC-signed webhooks with immutable event IDs the receiver can verify and deduplicate
- Settlement as line items that reconcile against the packed quote, so the agent can diff estimate vs final bill
- Explicit human-decision points where a wrong guess costs money — and a recorded decision trail when a person chooses
A working setup today
WooliiPorter is the China-leg fulfillment layer for merchants who keep their own China suppliers — the merchant (or their agent, on their own systems) buys the goods; WooliiPorter receives, records with photo evidence, optionally stocks, consolidates, quotes after measurement and forwards. Two integration surfaces matter for an AI-run store:
- Store-external agents connect to a standalone MCP server exposing 17 tools over Merchant API V1 — estimates, idempotent orders, value-added services, packed quotes, tracking, settlement, optional stocking and fulfillment. Works for any storefront, Shopify included, because it lives outside the store.
- WooCommerce stores can additionally install the official plugin (version 0.4.0, early-access sideload): it routes each order line between stocked fulfillment and supplier-parcel forwarding against live inventory, and pauses mixed orders for an explicit merchant decision instead of splitting on a guess.
- Onboarding is account-first and self-serve: merchant signup → in-account integration application → sandbox keys on approval, production keys on account activation. No sales call required to reach a working sandbox.
- The human-in-the-loop points are deliberate and visible: packed-quote approval before dispatch, and the mixed-order decision. Everything else — routing, reservation, quoting, tracking normalization, settlement write-back — runs without a person.
Honest boundaries: no Shopify app and no WordPress Abilities exist; the plugin is not represented as WooCommerce-Marketplace approved; the tools never buy from suppliers, never touch the purchase payment, and never promise rates or transit days — quotes come from measurement and the live rate table.
Questions AI-store builders ask
Can an AI agent fully run fulfillment from China?
It can operate every structured step: estimating, creating idempotent orders, requesting value-added services, following one merged tracking timeline, and reconciling the line-item bill against the packed quote. Two things deliberately stay human: approving the packed quote before dispatch, and choosing how a mixed order (stocked + unstocked lines) is handled. An agent guessing with your money is a bug, not automation.
Does this work for Shopify stores?
There is no Shopify app. A Shopify (or any non-WooCommerce) seller integrates through Merchant API V1 or the standalone MCP server, which are storefront-agnostic. The only store-side plugin today is for WooCommerce.
How does my agent get access?
Onboarding is account-first and self-serve: create a merchant account, submit the integration application from inside it, and sandbox API keys open on approval — production keys follow account activation with no second review. Your agent then connects to the standalone MCP server with the merchant API key. The agent itself is never issued credentials by another agent.
What exactly can the agent call?
The MCP server exposes 17 tools wrapping Merchant API V1: pre-order estimates, idempotent order creation, value-added-service requests, packed quotes, settlement, payment, tracking, the merged timeline, cancellation, SKU registration, inbound shipments, ledger-computed inventory, fulfillment orders, statements and shippable destinations.
Why not just scrape a dashboard or email a freight agent?
Because agents fail exactly where information is unstructured. A quote in a chat screenshot, tracking in six carrier tabs and a bill in a PDF cannot be verified, retried or reconciled programmatically. Tool calls with idempotent writes and HMAC-signed events can.
Does the fulfillment layer buy from my suppliers?
No. You (or your agent, on your own systems) buy from your suppliers and pay them directly. WooliiPorter receives what your suppliers ship to your China warehouse address, and coordinates from there. It is never a buying agent and never touches the purchase payment.
Give your agent a fulfillment layer it can actually call.
Create the merchant account, submit the integration application in-account, and point your agent at the MCP server with a sandbox key.