TL;DR
To choose an Amazon Ads MCP, ask five questions: who can do what, whether an engine keeps optimizing between chats, whether bulk changes are previewed, how many tokens each answer uses, and how long it takes.
Get every answer in writing, then test it on one real ad account. The same questions work for Amazon's own Ads MCP server, ads platform MCPs and self-hosted open-source servers. New to the term? Start with what an Amazon Ads MCP is.
Agencies and brands ask the same five questions but need different answers. An agency runs many client accounts with a mixed-seniority team; a brand runs fewer accounts, with more people outside the ads team asking for answers.
| Question | What an agency needs | What a brand needs |
|---|---|---|
| Permissions | Each client walled off; juniors propose, seniors apply | Finance and leadership read-only; the ads lead can change |
| Engine behind it | Client goals run between check-ins, so one manager covers more accounts | Targets held all day, not only when someone asks |
| Bulk changes | On event days, hourly reads and budget and bid changes across clients | Fix a budget running out at 2 p.m. on Prime Day in one step |
| Token use | Account managers on standard AI plans don't run out mid-week | Big catalogs stay inside the AI's working memory |
| Response time | A Monday review across a dozen accounts in one sitting | Answers fast enough to use in a meeting |
Permission control limits how much damage a mistake or a forgotten connection can do. For an agency, a connection that sees every client is a liability; for a brand, finance should never be one prompt away from changing a bid.
| Area | What to ask | Strong answer | Weak answer |
|---|---|---|---|
| Sign-in | Does each person sign in with their own login? | Per-person sign-in | One shared login or access key |
| Separation | Are read, edit, create and delete granted separately? | Each is its own permission | All-or-nothing access |
| Scope | Can a connection be limited to selected ad accounts? | Per connection | Per login only |
| Ceiling | Can a connection exceed the person's own role? | Never | Unclear |
| Lifetime | Does it expire, and who can revoke it? | Expiry you choose; revoke any time | Long-lived, no expiry |
Anthropic advises reviewing the permissions each custom connector requests. A typical split: analysts and leadership read-only, account managers with campaign edits on their own clients, and one senior lead who can create campaigns. The full order is in read-only vs read-write MCP.
An engine is optional, but it changes how the MCP works. An MCP with one keeps working toward your goals between conversations; an MCP without one acts only when someone asks. If every bid and budget move has to come from a prompt, your team still does the optimization, by typing instead of clicking.
Direct control is enough for small accounts, custom agents and teams that want to approve every bid. Engine-backed control suits large catalogs and agencies covering many clients.
Set target ACOS from each product's break-even ACOS (margin before ad spend), not a round number. A $40 product with $26 in cost and fees has a $14 margin, so break-even ACOS is 35%; a target of 25% to 28% leaves profit after ads. See AI-managed group levers and personality for the controls.
Bulk write on live data is where an Amazon Ads MCP pays for itself, especially on Prime Day, Prime Big Deal Days or Black Friday, when a hero campaign runs out of budget at 2 p.m. and category CPCs jump in the evening. An agency running a dozen clients can't catch that through exports and console clicks.
| Event-day moment | What you ask | What you apply in one bulk change |
|---|---|---|
| Budgets running out early | Which campaigns spent over [80]% of today's budget by [2 p.m.] with ACOS under target? | Raise those budgets [30]%, previewed old to new |
| Deal ASINs converting above plan | Which keywords on deal ASINs converted above [2]x their normal rate in the last [3] settled hours? | Raise their bids [15]%, capped at a maximum you set |
| Spend without sales | Which campaigns spent over $[200] since midnight with no orders and 20+ clicks? | Lower bids, previewed first; pause only the clear outliers |
| Event ends | List every budget and bid raised today. | Set them back in one change |
Any change you confirm goes live and spends real budget. Archiving on Amazon can't be undone, so pause instead. For the working flow, see bulk edit Amazon campaigns with AI.
Token use sets how many questions your team gets from the Claude or ChatGPT plans it already pays for, and the main driver is what the MCP hands back. An MCP that passes Amazon's raw reports straight to the model makes it pull reports, page through rows and calculate ACOS itself. Anthropic's engineers note that with direct tool calls, every intermediate result must pass through the model, and large results can overflow its working memory on big accounts.
MCP compatibility rarely separates providers anymore; token use does. For an agency, it is the difference between finishing the week's client questions on a standard plan and hitting the limit by Wednesday.
Response time decides whether an Amazon Ads MCP fits the working day. A question that takes 15 minutes won't be asked in a client call. Most latency comes from round trips: report requests, waiting for reports, paging and calculation inside the chat.
Amazon's own server is stronger for billing and first-party reach, a platform MCP for an engine to steer, and self-hosted for full control with a developer.
| Question | Amazon's own Ads MCP server | Ads platform MCP | Self-hosted open-source |
|---|---|---|---|
| 1. Permissions | Through API credentials and partner setup | Usually tied to platform logins and roles | Whatever you build |
| 2. Engine behind it | No group-level engine listed in Amazon's announcement | Sometimes, through the platform's automation | No |
| 3. Bulk write safely | Workflow tools; created campaigns need review | Varies; check that the server enforces previews | Whatever you build |
| 4. Token use | Depends on how much raw data each tool returns | Varies; ask for measured figures | Usually raw API output |
| 5. Response time | Depends on report round trips and paging | Varies; ask for measured timings | Depends on your build |
| Stronger when | You need first-party reach, billing or account settings | You want an engine to steer | You want full control and have a developer |
Many teams run both. See Amazon's own Ads MCP server vs a platform MCP.
Run it on one real ad account with a read-only connection, before anything is signed or any write access is granted.
Xnurta MCP is an ads platform MCP built on Xnurta's optimization engine. Its answers are below, limits included, so you can score it the same way as any other option.
| Question | Xnurta MCP |
|---|---|
| 1. Permissions | Read and write permissions granted separately; all or selected ad accounts; never more than the person's own Xnurta permissions; expiry after 30, 90 or 180 days or 1 year; revocation takes effect immediately |
| 2. Engine behind it | AI-managed groups: Xnurta's AI optimizes each group toward a goal and target ACOS with only the levers you turn on, plus a bid range and date-range schedules for events |
| 3. Bulk write | Old to new preview for budgets, bids and archiving, enforced by the server rather than the AI's prompt, then a read-back; operation logs by operator; live-campaign edits vary by ad type, with the most options for Sponsored Products; Sponsored Products campaigns created in bulk, all-or-nothing, optionally paused |
| 4. Token use | Designed to return compact, finished results instead of raw report rows, which keeps cost and token use down; measure it on your own account with step 2 above |
| 5. Speed | Designed for fewer round trips, because the analysis happens behind the server; time it with step 2 above |
| Where it's weaker | No billing or account-settings tools, which Amazon's own server lists |
To stop a change, decline the preview, pause the AI-managed group or revoke the connection in the Xnurta platform. Generally available from 27 Oct for Sponsored Ads. Write access is turned on per account by your CSM. Learn more on the Xnurta MCP page, or book a demo to run these five questions live.
Next question: Is it safe to connect Amazon Ads to ChatGPT or Claude?
FAQ
Plan on four weeks. Spend week one read-only, checking numbers against the console, then add campaign edits in week two and budgets in week three. Test automation or campaign creation in week four, and avoid starting a trial in the two weeks before Prime Day or Black Friday.
An Amazon Ads MCP has three cost parts: AI plan seats for each user, the tokens each answer uses, and server access. Token use is the part you can influence most, because an MCP that returns finished results instead of raw report rows uses fewer tokens per answer. Ask every provider whether MCP access is priced separately.
Coverage varies by server and by action. Most servers read all three ad types, but edits and campaign creation often have the most options for Sponsored Products. Ask each provider for a written table of reads, edits and creation by ad type.
Most changes can be reversed by editing back, and the change log shows what changed, when and by whom. Archiving on Amazon can't be undone, so pause instead and keep delete permissions off. Revoke the connection right away if a change looks wrong, then review every change it made that day.
It can, but it shouldn't. Give each client its own connection, limited to that client's ad accounts, with its own permissions and expiry. Then a mistake or a departing employee affects one client only, and each client's change log stays clean for client reviews.
Training depends on your AI plan, not the MCP. Anthropic and OpenAI say their business and API plans aren't trained on by default, while consumer plans differ. Use a business plan for client work and ask the MCP provider for its written data policy.
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.