Ecwid is an embedded store. It sits inside a site built on something else, so an agent working here has two sources of truth to keep straight: the Ecwid catalogue and whatever the host page says about the same product. We build agents that hold both, do catalogue work through /batch so bulk edits stay inside the rate limits, and wait for you before anything a buyer would see.
01Use
Most Ecwid stores are run by one or two people, so the work an agent takes off the pile matters more here, and so does the write it should never have made. Each of these is scoped narrowly enough to measure.
Fills in missing attributes, assigns categories, writes image alt text, and normalizes naming that drifted as the catalogue grew. Edits go out through /batch, so a catalogue-wide pass runs as batched writes rather than a request per product.
Answers pre-sale questions from the live catalogue plus your shipping and tax settings, on the storefront where the buyer already is rather than in a form they have to find.
Adds context to orders, drafts the follow-up message in your voice, and flags the orders that need a person to look at them before anything is sent.
02Wire
Ecwid's own API, the storefront it renders, and the site it is embedded in. The last one is the part people forget, and it is the one that drifts.
Read access is broad because answering well needs context. Write access is narrow, and every write is named before we build it.
| Agent | Reads | Writes | Human gate |
|---|---|---|---|
| Product data cleanup | Products, categories, product images, the host page | /products, /categories via /batch |
Price and inventory writes, always |
| Buyer questions | Products, shipping settings, tax settings | None. Read only | Low confidence escalates to you |
| Order follow-up | Orders, customers, the order's own history | /orders |
Every outbound message |
03Gate
There is no second reviewer to catch a bad write, and often no team dashboard anyone opens. So the default here is stricter than we would set on a larger store, and it stays that way until you tell us to open it up.
Every write to price, every write to inventory, and every message a customer would read. The agent prepares the change and stops. Nothing in that list runs unattended, whatever the agent's confidence is.
By email, with the trace attached: what triggered the run, what the agent read, the field it wants to change and the value it would replace. You approve or reject from the mail, because a dashboard nobody checks is not a gate.
When the Ecwid catalogue and the host page disagree about the same product, the agent reports the difference rather than picking a winner. Reconciling the two is a decision about your store, so it is yours to make.
Each agent holds only the OAuth scopes its task needs, so a cleanup agent cannot reach orders. Every write is recorded with the value it replaced, which is what makes a mistake reversible instead of archaeological.
Related: retrieval and data for grounding answers in your own product content, and automation and workflows for the triggers and approval steps around a run.
04Asked
It changes what the agent has to reconcile. The Ecwid catalogue and the host page around it are two records of the same product, and they drift apart. We treat the Ecwid REST API v3 record as the one the agent writes to, then read the host page and report where the two disagree instead of quietly overwriting either side. Agent work on the host page itself is a separate build.
No. We ask for the OAuth scopes one task needs and nothing else. A buyer question agent reads products and your shipping settings and can write nothing at all. A cleanup agent writes to products and categories but never sees orders or customers. The scopes are written down in the build, so you can see what was granted and withdraw it.
The run trace records the field it changed and the value it replaced, so a wrong edit is visible and reversible. Price and inventory never move without your approval, so the worst unattended case is a description or a category you can correct. That example then goes into the eval suite, so the same mistake is caught before the next change ships.
Name the task you keep putting off in the Ecwid control panel, and the site the store is embedded in. We will tell you what an agent would touch, what would wait for you, and what it would cost to run.
