Follow one Amazon answer through Claude, an MCP server and the authorized source, including the less obvious problem of old cached extracts.
The number arrived. Where did it come from?
Picture a reply saying last week's ordered product sales were $42,000. Before repeating it in a meeting, ask for the account, marketplace, source report, sales field, date boundaries, reporting time zone, retrieval time and row grain. The number could come from a cached weekly ASIN extract rather than a fresh request. That is not automatically wrong: a frozen report is useful for a retrospective. It is wrong to present that extract as a live account total or to hide its age during an inventory decision.
1stPage has a customer-facing Connect MCP page with the remote server URL. Add that URL as a custom connector in a supported Claude client, then sign in through 1stPage’s OAuth screen. Ask list_accounts which accounts your user can see before calling an analytics tool. A tool call names an account and marketplace; OAuth approval itself is not a per-account selection screen, so review user access before connecting.
In a separately built workflow, the route is Claude client → MCP tool/server → authorized data service → Amazon source or stored extract. Model Context Protocol describes how the client finds and calls tools; it does not grant permission to Seller Central, Vendor Central or Amazon Ads. The service behind the tool must already have the right authorization for the particular operation.
A successful tool call proves that a call completed, not that the returned figure has the requested scope. Keep the source metadata in the answer, where a reviewer can compare it with the original report.
Name the tools available in this session. For every Amazon figure, identify the account, marketplace, report, metric, grain, reporting dates, time zone and retrieval time. Distinguish a cached extract from a fresh request. If the authorized source cannot provide the requested figure, say it is unavailable rather than estimating it.
A failed grant is not a zero
Amazon's SP-API authorization and roles belong to the selling partner and registered application. An operation lacking its required role can return 403. Data Kiosk also checks access to requested fields when a query is created. A different prompt cannot repair a missing role. Ask the account owner to review the approved application and requested scope; never translate a denied query into zero sales. An expired grant, a processing report and a genuinely empty report deserve three distinct statuses in the response.
A seller authorization does not provide a vendor account, and neither seller nor vendor report access implies Amazon Ads API access. Multiple connected accounts add another trap: an unlabeled blended total can look credible while crossing marketplace and currency boundaries. Require an explicit account selection for each query. Analytical tools need read operations, not a broad write path. Keep OAuth and refresh tokens in the service's credential store, outside prompts, chat transcripts, pasted logs and spreadsheets. Record who authorized each source and who can revoke it.
Disconnect it before calling it connected
Run one narrow read on an approved report and compare a subtotal with the source export for exactly the same dates. Then test an expired or revoked grant in a safe environment. A properly bounded connection returns an authorization failure rather than quietly recycling last month's number. Inspect the retention rule for extracts separately: revoking a live grant does not erase earlier files or a conversation that quoted them. The team also needs a way to disable the server, remove an individual tool and review access history.
That pair of tests is more revealing than a screenshot of a populated tool list. Record the expected report completion state and the response when no rows exist. An unavailable report must not inherit the last successful value. For a one-off question, an approved export stripped to the needed fields can avoid a standing connection altogether; recurring analysis needs a named owner, role, retention rule and reconciliation check. This is a test plan for a possible deployment, not a statement that 1stPage provides a particular MCP server.
Which Claude surface did you test?
Anthropic documents remote custom connectors in Claude.ai. Claude Free currently permits one; Team and Enterprise owners add a connector before members can authenticate to it. A connected tool still needs the right 1stPage account and Amazon report behind the question. Specify the actual client, plan, server transport and tool name before publishing setup instructions. Check authentication with a real approved read, not just a tool-list display. A server can expose a tool while its downstream Amazon grant is absent. Keep client permissions and source permissions in separate boxes on any architecture drawing. The first determines whether Claude may call the named tool. The second determines whether the service may retrieve the report. A security review should inspect both and trace one returned field back through them. This matters most when a team changes clients but reuses a service account: the visible interface changed, while the Amazon permission boundary did not.
Do not confuse this with Amazon's plugin
Amazon announced a Selling Partner plugin in Amazon Quick and in beta with Claude, initially for sellers in U.S. stores. Amazon describes Seller Assistant reasoning, recommendations and seller-approved actions with boundaries and audit trails. That official route exists alongside independently built MCP analysis. The announcement does not promise eligibility for every account, access to every SP-API report, vendor data, Ads data or Amazon Marketing Cloud results.
For a seller with access, inspect what the plugin actually shows and which actions need approval. For a separate analysis service, inspect the named reports and read-only tools its own authorizations permit. Do not equate Amazon's promotional Quick Plus subscription with a Claude subscription. The practical choice is the specific question and available source, not a blanket comparison of assistants. Ask which route supplied the named report before comparing answers.
Sources and boundaries
Check the current permissions and report definitions before acting on an answer. Example prompts are questions, not live account results.