Explainer
TL;DR
It is safe to let an AI agent change Amazon ad bids and budgets when you control what it can see, what it can change and when it has to ask. With Xnurta MCP, the AI works under your own Xnurta login, each kind of write has its own permission scope, and changes to money, bids or archiving show a preview and wait for your yes. This guide is for marketers and agency teams who already use Claude or ChatGPT for ad analysis, and for whoever has to approve the connection before it touches a live account.
It is safe when four things hold, and with Xnurta MCP you can check each one yourself. The AI works under your login. Reading never changes anything. Writing needs its own permission, and anything that moves money waits for you. Afterward you can see what changed.
Xnurta MCP connects Claude, ChatGPT and other MCP clients to your Xnurta data. That sentence should make a careful person nervous, so the rest of this guide is the proof: what the AI can reach, what it can touch, where it stops and asks, and where the trail lives. A preview only protects you if somebody reads it. The controls give you the chance to.
Yours. With OAuth, you sign in to Xnurta in your browser with your own login, review the permissions and ad accounts it asks for, and approve. The AI then sees only the ad accounts and permissions that login carries, main accounts and sub-accounts included. There is no token to create, copy or store.
Teams add one wrinkle. On Claude Team or Enterprise, an Owner or Primary Owner adds the Xnurta connector once for the organization, as Anthropic's help center describes. Each member then connects on their own and authorizes with their own Xnurta account. Adding the connector for the organization does not widen anyone's access.
You do not have to take the scope on faith. Ask the assistant.
Prompt: show me what you can reach
Call get_user_authorized_context and list the ad accounts I'm authorized to access.
The get_user_authorized_context tool returns your user ID and the ad accounts your login is authorized for. If an account is missing from that list, the AI cannot see it. If one shows up that should not, find out why before you go further. Connect Claude or ChatGPT to Xnurta MCP has the connection steps.
No. Queries are read only. Xnurta MCP has three query tools, for ads performance, entity metadata and the operation log, and none of them modifies your account. You ask in plain language and the assistant picks the tool.
The answer comes back into your conversation with your AI assistant. That is the point where Xnurta's controls end and your AI provider's begin, which is the next question.
Three parties each control a different layer: Xnurta decides what the AI can reach, your AI provider handles the conversation, and your team decides what gets confirmed.
| Layer | What it governs | Who sets the rules |
|---|---|---|
| Your Xnurta login and permissions | Which ad accounts the AI can reach, and whether write access is on | Your Xnurta account permissions, and your CSM for write access |
| Your AI assistant (Claude, ChatGPT or another client) | The conversation, including the data the tools return | Your own AI plan and settings |
| Your team | Every preview before a change | Whoever reads it and says yes |
On model training, retention and similar questions, the answer sits in the middle row. How the AI provider handles conversation data follows your own AI plan and settings. We do not summarize those policies here, because they belong to the provider and they change. Read the data controls on your Claude or ChatGPT plan, and read your agreement if you work with client data.
The AI can make three kinds of write, each with its own permission scope, and anything involving money, bids or archiving needs your confirmation. A scope that lets the AI edit ads does not necessarily let it create campaigns.
| Kind of write | What it covers | Where it stops |
|---|---|---|
| Managed group management | Create, edit and delete AI managed groups: name, AI status, personality, goal, target ACOS, total budget, campaigns in the group, action space, schedule, template | Ask for the objects and settings it plans to change, then confirm. Deleting a managed group archives it, and archiving cannot be undone. |
| Ad-entity editing | Campaign state, budget and bidding strategy, products, negatives, placement and audience bid adjustments, keyword and product targets with state and bid | Money, bids and archiving return a preview first and submit only after you confirm. |
| SP campaign creation | Sponsored Products campaigns with ad groups, products, targeting, negatives, placement bid adjustments and an optional AMC audience | Creation is atomic: if one campaign fails validation, none are created. Ask to see the full structure before it submits. |
Some settings are out of the AI's reach entirely. Brand, non-brand and competitor mode, and every word-list feature, can be changed only in the Xnurta platform.
Write access is also not something that switches on by accident.
The access rule
Generally available from 27 Oct for Sponsored Ads. Write access is turned on per account by your CSM.
After your CSM turns it on, you remove the Xnurta connector and add it again, or create a new token with the scope you need. Signing in again on the old connector keeps the old permissions. Then you update your Xnurta skills.
You make the assistant show its work first. For money, bids and archiving it does that on its own: a preview of which objects will change, from what value to what value, and nothing is submitted until you confirm. For managed groups, ask for the same list yourself.
Prompt: preview a budget change (needs ad-entity editing)
In ad account [account name], raise the daily budget on [campaign A], [campaign B] and [campaign C] by [10]%. Show each campaign with its current and new budget first. Do not submit until I confirm.
Prompt: build campaigns paused (needs campaign creation)
Create a paused SP auto campaign for these ASINs in ad account [account name]: [ASIN list], with a daily budget of [amount]. Show me the full structure before you submit it.
The second prompt matters for a reason. Campaigns cannot be deleted, only archived, and archiving is irreversible. Create them paused if you want to review before they spend. And pause a campaign when you only want it to stop, never archive it.
You ask the operation log. It holds human and AI actions on campaigns, ad groups, targets, bids, budgets and managed groups. You can filter by operator (AI, human, automation rule, or one person's email), by action type, by entity type, by ad type and by Amazon entity ID, and the log reaches back about 15 months.
You do not need the filter names. Say what you want.
Prompt: what did AI do in my account this week
Which automatic bid changes and negative-targeting actions did AI perform in ad account [account name] over the last 7 days, and why?
Asking the log is a query, so it changes nothing. Who changed my Amazon campaign? walks through the filters one question at a time.
You treat it like a password, or you skip it. OAuth needs no token at all, so use it wherever your client supports it. Claude desktop, Cowork and claude.ai have no field for a token, so they use OAuth through the custom connector.
For clients that do take a token, you create it in Xnurta under MCP & Skill, click New token and choose the scope. Pick the narrowest scope that covers your first use. Copy the token once and store it securely, because you cannot view it again after you leave the page. In ChatGPT Codex, point the configuration at an environment variable (the Xnurta guide uses XNURTA_TOKEN) so the token never sits in the config file.
You start where nothing can change, then widen one step at a time.
get_user_authorized_context check above. Confirm the ad accounts are the ones you expect.FAQ
Not for money, bids or archiving. The assistant returns a preview of the objects and the old and new values, and submits only after you confirm.
Yes. Queries are read only, and all three kinds of write need write access, which your CSM turns on per account.
No. With OAuth it uses your own Xnurta login and sees only the ad accounts and permissions that login carries, main accounts and sub-accounts included.
That follows your own AI plan and settings with the provider. Check the data controls on your plan with the provider.
Ask the operation log, which holds human and AI actions on campaigns, ad groups, targets, bids, budgets and managed groups. You can filter by operator, action type, entity type and more.
Keep going
Xnurta MCP
Xnurta MCP is for Xnurta customers only. Create your MCP connection directly in the Xnurta platform. New to Xnurta? Book a demo and we'll walk through your account with you.
