Amazon Ads MCP for Authors: Connect Your AI Assistant

An author can connect an MCP-compatible AI assistant to Publisher Ads with an API key, then ask questions about the Amazon advertising accounts available to that member. Start with a narrow read-only request, confirm the account and date range, and treat every permitted write as a possible live change to an ad.

What does an Amazon Ads MCP connection do?

MCP, or Model Context Protocol, gives a compatible client a structured way to call tools supplied by another application. In this case, the client can work with the account-scoped advertising information and actions made available by Publisher Ads.

That is different from pasting a report into a chat. A connected client may be able to list books and campaigns, inspect targets and search terms, compare performance over an exact period, review change history, or use permitted campaign-management tools. Availability depends on the member account, the client and the active permissions.

The connection does not turn a general assistant into an all-knowing ads manager. It still needs a clear account, marketplace, currency, book or campaign identifier, and time period. It also cannot overrule Amazon's eligibility or review decisions.

1. Check that your AI client supports the connection

Your client must support Streamable HTTP servers and bearer-token authentication. Claude Code and Codex are examples of clients an author may use, but the exact setup screen and labels can change between client versions.

Use the client's normal server-connection area rather than copying an old command from a forum post. You will need two pieces of information:

1. The server endpoint: https://publisherads.ai/mcp
2. A bearer token generated as an API key in Publisher Ads Settings

The AI client's own subscription, API allowance and usage terms remain separate. The connection does not include reasoning usage from the client.

2. Generate a dedicated API key in Settings

Open Settings in your Publisher Ads account and use the API-key section to generate a key. Give the connection its own key when practical, rather than reusing one token across several devices or clients. Separate keys make it easier to revoke one connection without disturbing another.

Treat the key like a password. Put it only into the client that will make the connection, do not paste it into a shared document, and do not include it in a support screenshot. If you think it has been exposed, revoke it and generate a replacement.

Publisher Ads is campaign software for self-published authors and publishers. Its MCP connection uses the same member account context, so it is worth checking that the Amazon advertising account you expect is already connected before troubleshooting the AI client.

3. Add the server without relying on brittle commands

In the client's MCP or integrations settings, create a Streamable HTTP connection. Enter the endpoint exactly as shown above and set the generated API key as the bearer token.

The names of these fields vary. One client may say “HTTP server,” another may say “remote MCP,” and another may ask for an authorization header. The important parts are the transport, endpoint and bearer token, not the wording of the setup form.

After saving, ask the client to show the tools or capabilities it can currently access. Do not assume that every account receives an identical tool list. Feature access and account configuration can differ, and some operations may be unavailable even when the basic connection works.

4. Make the first request read-only and specific

A good first request proves identity and scope without changing anything. For example:

List the Amazon advertising accounts I can access. Do not make changes.

Then narrow the request:

For the US account named in the previous result, list enabled Sponsored Products campaigns for book ASIN B0EXAMPLE1 from October 1 through October 14, 2026. Show spend and sales in US dollars. Do not make changes.

The ASIN and dates in that example are hypothetical. The useful pattern is real: state the account, marketplace, status, identifier, dates, currency and read-only boundary.

If the answer seems wrong, first check scope. Authors often have similarly named books, multiple formats, several marketplaces, or more than one Amazon advertising account. A clean identifier prevents the assistant from guessing which “Book 1” you meant.

5. Ask the assistant to explain the evidence

Before acting, ask for the rows or facts behind a recommendation. A useful request might be:

Find search terms with spend in this period and group them into terms with attributed orders, terms without attributed orders, and terms with incomplete or unclear data. Cite the campaign and ad group for each. Do not change bids or negatives.

This creates a reviewable result. It also keeps unknown data separate from zero. Missing placement data, incomplete dates or a recently launched campaign should not silently become evidence that something performed badly.

For a broader overview of where an AI connection fits alongside research and direct campaign tools, see the guide to Amazon ads tools for authors.

6. Read the active permissions before any write

Do not rely on a remembered setup or an old screenshot. Ask the client to report the active tools and permissions for the current connection. Then name the exact action you are considering.

This matters because permitted MCP write actions can change live advertising. They do not all wait for a universal website approval card. Depending on the active configuration, the connection may be able to rename a campaign, change its status, adjust a budget or bid, manage certain negatives, or perform another supported action.

That makes a vague instruction such as “fix my ads” too broad. Prefer a bounded instruction:

On campaign 123456 only, preview the effect of changing the daily budget from the current value to the hypothetical value of $8. Do not apply it. Report the current marketplace and currency first.

When a preview is available, use it. When it is not, ask the assistant to restate the target, present value and intended new value before proceeding.

7. Apply one small change and inspect the result

Your first write should be easy to recognize and easy to reverse. A campaign rename is often clearer for a connection test than a wide set of bid edits, provided renaming is available and you have chosen a non-critical campaign.

Use a request that contains one account, one campaign and one action. For example:

Rename campaign 123456 from its current name to “US Test - Exact - October 2026.” Apply only that change. Then read the campaign again and show the recorded result.

The campaign details above are hypothetical. The safety pattern is the point: identify, restate, change, then verify. Publisher Ads records applied changes in history, so you can also ask the client to retrieve the relevant change record.

For a budget, bid, status or targeting action, add another check: confirm that the campaign type supports the requested operation. Fixed-bid stack layers can have special bid guards, and a quiet campaign is not automatically one that needs a higher bid.

8. Build a repeatable request template

A short request template reduces ambiguity when you return each week:

1. Scope: account, country, book, campaign and format.
2. Period: exact start and end dates, plus the currency.
3. Evidence: the fields you want listed before a recommendation.
4. Boundary: read-only, preview only, or one explicitly permitted change.
5. Verification: read back the result and show its history record.

For example, “Review my campaigns” leaves nearly every important choice unstated. “Read the UK campaigns for ASIN B0EXAMPLE2 from October 1–14 in GBP, flag targets with spend but no attributed orders, and make no changes” gives the client a testable job.

If you later automate a repeated review, keep the same boundaries. A recurring audit should not quietly inherit permission to edit every campaign simply because one earlier request allowed a write.

9. Revoke access when the connection is no longer needed

Revoke the API key in Settings when you retire a client, lose control of a device, or finish a temporary workflow. Then confirm that the old connection can no longer list account data.

If you reconnect later, create a new key and repeat the read-only identity check. That small step is faster than investigating an apparent data problem caused by a client pointing at the wrong account or using an expired credential.

What is a sensible first Amazon Ads request?

Start with inventory, not optimization. List the accessible accounts, select one marketplace, list a small group of campaigns for one book, and request exact dates and currency. Once those identifiers are correct, move to search terms, targets, budgets or change history.

The best early outcome is not a dramatic campaign change. It is confidence that the assistant is reading the right account, interpreting the requested period correctly and respecting the boundary you gave it.