
An e-commerce integration looks like a content integration until something is written. A wrong product description is an edit; a wrong order status is a customer receiving the wrong thing, or a refund that never happens. That asymmetry should decide the permissions.
Two servers cover the mainstream stacks: Shopify and WooCommerce.
The short answer
Connect read access for support and merchandising questions. Keep order writes, refunds, and stock changes behind a human approval. The read path answers most questions and carries almost no risk.
What the read path does
| Question | Read path |
|---|---|
| Where is this order? | Order and fulfilment status |
| Do we have this in stock? | Inventory by variant |
| Why was this returned? | Return records and notes |
| What does the product page claim? | Product description and attributes |
Each is a single lookup with a factual answer, which is exactly the shape an agent handles well. None requires a write.
Shopify: the API is the store
Shopify exposes admin data — products, variants, orders, inventory — through the Admin API. A read-only token scoped to one store is a small, useful integration: support can ask questions in the tool they already use.
The tempting next step is order management: updating status, issuing refunds. That is a different product, and it needs a review surface, because a refund is money.
WooCommerce: the same shape, different ownership
WooCommerce runs on WordPress, which means the store owner also owns the hosting and the database. That shifts the risk: an integration with write access is also an integration with the site itself.
A read-only consumer key restricts what the agent can reach, and it keeps the blast radius to the store data rather than the installation.
The write decisions, ranked
Product description edits are low risk and need review for accuracy. Inventory changes are medium risk and should be automated only from a system of record. Order status and refunds are high risk and need a person. Customer data changes are highest and usually should not be an agent capability at all.
A support workflow that works
An agent reads the order, the ticket, and the product page, then drafts a reply with the facts. A person checks the answer and sends. The agent never changes the order; if a change is needed, it produces the exact action for the person to take in the admin.
That last sentence is the design: the agent proposes an action with its arguments, and the person executes it. It keeps the audit trail human and the agent useful.
Frequently asked questions
Can an agent answer 'where is my order'? Yes, with read access, and that is the highest-volume support question in most stores.
What about inventory sync? Synchronise from the system of record, not through an agent. See automation tools.
Which store first? The one with the support queue. See MCP servers.


