Connect Amazon Seller Central to ChatGPT
Give ChatGPT read access to your own Amazon numbers and the ability to draft changes you approve. Figures come from the settlement events Amazon posted to your account, not from a fee estimator, and every answer carries its basis and its gaps.
Set it up in four steps
- Create an AgentXray account.
- Connect Seller Central — one Amazon login. AgentXray reads orders, fees and settlements over SP-API.
- Copy the connector URL:
https://agentxray.ai/api/mcp - In ChatGPT, open Settings and find Apps. Developer mode is its own row inside Apps; opening it jumps to Account security and carries a high risk label. Create the connector with the + at the top right of the app list; the dialog makes you tick a risk acknowledgement before it will save.
It is not called Connectors. Most guides still say it is — ours did too, until we walked it on 2026-09-03. The old URL /#settings/Connectors silently redirects to /#settings/Plugins, which is why a stale name went on looking plausible for years. Setup was confirmed working on a free ChatGPT account in the same session.
What you can ask
- Profit and fees — what a unit actually earned after the referral, FBA and storage amounts Amazon posted; break-even price; net contribution by SKU.
- Advertising — spend, search terms, placements and negatives, judged against that product's measured contribution rather than an ACoS target.
- Inventory and Buy Box — stock age, restock timing, and who holds the Buy Box right now.
- Listings — what is live, what is suppressed, and which issues Amazon is reporting against it.
Ask what_i_cannot_tell_you first. It reports the measured limits of your own data — which marketplaces are connected, how far the settlement history reaches, what is missing — before you draw a conclusion from a number.
What it can draft, and who approves it
Of the 32 tools, 19 only read. The other 13 write nothing to Amazon: they file a draft — a bid, a budget, a negative keyword, a campaign state, a listing edit, a shipment — into an approval queue you open in your browser.
The separation is structural, not a policy we promise to follow. The connector token that files the draft cannot authenticate against the approval endpoint; it is refused with a 401. An assistant that has been talked into approving its own draft still cannot, because the credential it holds is not accepted there.
What we never see
- Buyer names and addresses do not reach us at all. AgentXray holds no Amazon PII role, so that data is never sent. This is a smaller claim than logging access to it, and a stronger one.
- Row-level security, not a filtered query. The request path connects as a database role that cannot bypass it, so a session that has not declared whose data it is reads nothing rather than everyone’s. A live cross-account attempt on 2026-08-29 could not write across the boundary.
- Least privilege. The application role holds no superuser rights and cannot alter the schema.
Two things we are not claiming, because we have not measured them: an annual third-party penetration test, and encryption-at-rest attestation from the hosting provider. Both are outstanding, and a trust page that listed them anyway would be the first thing worth distrusting.