Explainer
TL;DR
Agencies should ask four things before connecting an AI assistant to a client's ad account: whose login it uses, what it can read, what it can change, and where the record lives. Xnurta MCP answers all four. With OAuth the assistant signs in with each person's own Xnurta login, reading never changes anything, and changes to money, bids or archiving wait for a preview and your confirmation.
With OAuth, the AI uses the login of the person who connects it, and nothing else. Xnurta MCP connects Claude, ChatGPT and other MCP clients to your Xnurta data. The browser opens the Xnurta authorization page, shows the permissions and ad accounts being requested, and waits for the person to approve. Nobody creates, copies or stores a token.
Whatever that login can open in Xnurta is what the assistant can see. Main accounts and sub-accounts both count. Not one client more.
| Question | Answer |
|---|---|
| 1. Does the AI sign in with a shared agency login? | Not with OAuth. It uses the connecting person's own Xnurta login, and the assistant sees only the ad accounts and permissions that login carries. |
| 2. Are sub-accounts included? | Yes. Main accounts and sub-accounts follow the same rule: the assistant can query them if the person's Xnurta login can access them. |
| 3. Can a teammate see a client account their login cannot open? | No. With OAuth, each member authorizes with their own Xnurta account and sees only the ad accounts and permissions it carries. |
| 4. How do I check which ad accounts the assistant sees? | Ask it to call get_user_authorized_context. It returns your user ID and your authorized ad accounts. |
| 5. Does our Claude Team or Enterprise organization need an admin? | Once. An Owner or Primary Owner adds the connector for the organization, then each member connects with their own login. Adding it for the organization widens nobody's access. |
| 6. Do we have to manage tokens? | Not in Claude desktop, Cowork or claude.ai. They connect through a custom connector with browser sign-in and have no field for a token. Clients that send one take a token created under MCP & Skill, then New token, shown once. |
It reads performance, settings and the operation log, and reading changes nothing. Three query tools cover those, and all of them are read only.
A Monday check across clients is one sentence. Ask the assistant to list the ad accounts you are authorized to access and show how each performed yesterday. Then name the client account in every follow-up, so an answer never lands on the wrong brand.
| Question | Answer |
|---|---|
| 7. Does asking a question change anything in a client account? | No. Queries on ad performance, entity settings and the operation log are read only. |
| 8. What can the assistant read? | Ad performance, which can also be broken out by audience for campaigns. Entity metadata such as campaigns, ad groups, targets and managed groups. And the operation log. |
| 9. How far back does it go? | About 15 months of performance data and logs, across 22 marketplaces. |
| 10. Is the data live? | Not to the second. Freshness follows the Xnurta platform's update cadence. Data is daily by default, with hourly data for SP, SB and SD campaigns and SP keywords. |
It can change a client's account only where write access is on, and the risky changes wait for you. Writing comes in three kinds, and each kind has its own permission scope. That matters for agencies: a scope that edits ads does not necessarily create campaigns. Read-only vs read-write MCP for Amazon advertisers explains which access to start with.
The access rule
Generally available from 27 Oct for Sponsored Ads. Write access is turned on per account by your CSM.
| Question | Answer |
|---|---|
| 11. Can the AI change a client's campaigns? | Only where write access is on for that ad account. Reading needs no write access. |
| 12. What kinds of change are there? | Three, each with its own permission scope: managed group management, ad-entity editing (budgets, bids, state, keywords, negatives) and Sponsored Products campaign creation. |
| 13. What needs my confirmation? | Anything involving money, bids or archiving. The assistant returns a preview first, showing which objects and from which value to which value, and submits only after you confirm. |
| 14. Can a change be undone? | Archiving cannot. It is irreversible, deleting a managed group archives it, and campaigns cannot be deleted. Pause a campaign if you only want it to stop. |
| 15. What if one campaign in a bulk creation fails? | Creation is atomic. If any campaign in the request fails validation, none of them are created, so a client account is never left half built. |
| 16. Can the AI build a campaign without it spending? | Yes. Campaigns are created in the state you ask for, so ask for paused and review the full structure before it submits. |
| 17. What can the AI not change? | Brand, non-brand and competitor mode, and every word-list feature, stay in the Xnurta platform. They are not writable through MCP. |
You ask the operation log, in plain language, and you correct a change by asking for the value you want. The log holds human and AI actions on campaigns, ad groups, targets, bids, budgets and managed groups. For an agency, that turns "who changed this?" into a lookup. Who changed my Amazon campaign? has the prompts.
| Question | Answer |
|---|---|
| 18. Where do I see what changed? | In the operation log, which reaches back about 15 months. Ask for one campaign, one ad account or one time window. |
| 19. Can I filter the log by who acted? | Yes. Filter by operator: AI, human, automation rule, or one person's email. You can also filter by action type, entity type and the Amazon entity ID. |
| 20. How do I override a change I don't like? | Tell the assistant the value you want. For money, bids and archiving it previews the old and new values and waits for your confirmation. |
Connect people one at a time, install the skills as a separate step, and test on one client first. Anthropic's help center covers who can add connectors in Get started with custom connectors using remote MCP. The endpoint is https://mcp.xnurta.com/mcp, and Connect Claude or ChatGPT to Xnurta MCP walks through each client.
Adding a connector never installs skills. Skills teach the assistant the routine for Xnurta tasks, so ask it to read the Xnurta MCP skills repository on GitHub and install them in a new chat. Two official skills fit repeat client reporting. Weekly Ads Report produces week-on-week KPIs, 7-day trends, anomalies, AI and automation activity, top movers and next-week actions. Monthly Ads Report adds month-on-month and year-on-year baselines.
Your own routines can become skills too. You keep them private or share them with your organization, and if a member leaves, the main account can hand over their skills so the playbook stays. How an SOP becomes something an AI assistant can run and Client reporting with MCP: from question to slide go further.
Prompt: first check, read only
List all ad accounts I am authorized to access. How did each one perform yesterday? Identify the five issues that need the most attention and support each one with data.
Prompt: who touched a client budget
In ad account [client account name], who changed campaign [campaign name or ID]'s budget, and when? List all human actions from the last [30] days.
It covers access, change control and the log. It does not cover where data is stored or whether any of it trains a model. Is it safe to connect Amazon Ads to ChatGPT or Claude? picks up that thread, and anything your client contracts demand beyond it belongs in a written question to your Xnurta contact. Ask the same four questions of every vendor you test: whose login, what can it read, what can it change, where is the record.
FAQ
Not with OAuth. It uses the connecting person's own Xnurta login, so it sees only the ad accounts and permissions that login carries, main accounts and sub-accounts included.
Only where write access is on for that account. Generally available from 27 Oct for Sponsored Ads. Write access is turned on per account by your CSM.
Not on money, bids or archiving. It shows a preview of the objects and the old and new values, and submits only after you confirm.
Yes. The operation log holds human and AI actions, and you can filter by operator, including one person's email, or by the Amazon entity ID.
Not in Claude desktop, Cowork or claude.ai, which connect through a custom connector and browser sign-in. Tokens are for clients that can send one.
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.
