# Stay in control (https://sellersheetai.com/docs/mcp-api/control)

> Your agent works for you. Labels on every tool, write access per store, approval in your client and trust you widen step by step keep it that way.



## Every tool says what it does [#every-tool-says-what-it-does]

Each tool carries one label, and your AI client reads the same label:

| Label                                | Meaning                                                                                                       |
| ------------------------------------ | ------------------------------------------------------------------------------------------------------------- |
| <AccessBadge access="read" />        | Reads only — changes nothing.                                                                                 |
| <AccessBadge access="write" />       | Creates or changes something.                                                                                 |
| <AccessBadge access="destructive" /> | Changes or removes something that cannot simply be undone — a live listing, a confirmed shipment plan, a bid. |

No tool both reads and writes, so a "look" request can never change anything by accident. The full rules: [Read, write & destructive](/docs/mcp-api/access).

## Turn on tool approval [#turn-on-tool-approval]

Most clients can ask before every tool call. Keep it on for anything that writes. Many let you allow read-only tools once and keep asking for the rest — a good place to start.

## Write access is per store [#write-access-is-per-store]

Reading always works on a store you can see. Changing it needs write access **on that store**, or the call is refused with 403 before anything reaches Amazon:

* **Your keys** carry Seller API and Ads API access per store — none, read, or read and write ([Keys & store access](/docs/mcp-api/keys)).
* **Your team**: the owner or an admin of the workspace sets each member's Seller API and Ads API access per store, and one level per noon account, under **Settings → Members**.
* A key never reaches beyond what its holder can reach, so a read-only member's key is read-only too.

Report and export *requests* — asking Amazon to prepare a report — are reads: they change nothing in your account.

## Trust grows over time [#trust-grows-over-time]

Start by approving everything. Ask your agent to write what it intends into a Google Sheet first — the bids it would change, the listings it would submit — and approve there; the sheet is also your team's record. Once a kind of work has gone right many times, let it run on its own: low-risk reads and reports first, changes later. Your agent never picks an option for you — a placement, a transport offer — unless you give it the rule.

## Revoke any time [#revoke-any-time]

* **An AI client or a key:** **Settings → API & MCP** lists every key, including the one each browser sign-in created. Turn it off or delete it and it stops working at once.
* **SellerSheet's access to Amazon:** Seller Central → Manage Your Apps. Your stores then stop syncing and refuse calls until you reauthorize.

See also: [Data we never see](/docs/start/privacy) · [Troubleshooting](/help/troubleshooting/ai).