On September 23, 2026, at Amazon Accelerate in Seattle, Amazon announced that sellers can connect Seller Central to Anthropic's Claude. The same day, Amazon's plugin landed in Anthropic's official plugin directory. It is free, it is in beta for sellers in Amazon's US stores, and the first question every seller asks is the right one: what does it actually do?
This is a plain description of the plugin from its own source code and documentation, which Amazon published openly, followed by what it deliberately leaves out, how it sits next to the separate Amazon Ads MCP server that has been in open beta since February, and where a third-party connection like ours still earns its place. Everything here is dated and sourced at the end. We build a competing connector, so read the last third with that in mind. The first two thirds are just the facts.
The short version
- What it is: a remote MCP server run by Amazon, plus a set of skills that teach Claude how to use it. You authorize it with your Seller Central login. Claude gets only the tools your Seller Central user role allows.
- What it covers: FBA inventory, sales and traffic analytics, stockout risk and restock timing, FBA inbound shipments end to end, listing issues, buyability, search content, a compliance pre-flight, support case preparation, and a documentation assistant for developers.
- How it writes: every change is drafted and previewed, one at a time, and nothing goes live until you approve it in so many words.
- What it does not cover: advertising, orders beyond aggregate metrics, settlements, fees, tax reports, cost of goods, profit, Vendor Central, more than one marketplace per session, and any pricing advice at all.
- Where it runs: Claude, in beta for US stores, and Amazon's own Quick assistant. More assistants are promised. ChatGPT is not supported at launch.
What was announced, and by whom
Amazon's announcement came from the Accelerate keynote. Mary Beth Westmoreland, Amazon's VP of Worldwide Selling Partner Experience, framed the goal as bringing Seller Central to wherever sellers already work, so they need not log into it. Anthropic's field CTO, Eric Burns, described the plugin as bringing Seller Assistant's insights into a seller's Claude account alongside their existing accounting and supplier connections. Amazon says connecting takes about a minute and needs no code.
Three related things shipped together and are easy to blur:
- The Amazon Selling Partner plugin for Claude, the subject of this post. Beta, US stores, free.
- Seller Assistant upgrades inside Seller Central: persistent memory of your pricing patterns, inventory cycles and goals, and always-on workflows that can either surface recommendations or take approved actions in the background. Amazon says that memory follows you across Seller Central, Quick and Claude.
- Amazon Quick Plus, free for 12 months to every primary account holder through December 31, 2026, with two co-workers included.
How it works under the hood
Amazon published the plugin's source openly, under an Apache license, in a repository called the Selling Partner agentic toolkit. That makes it unusually easy to say precisely what it is.
- One remote MCP server. The bundled configuration points Claude at an Amazon-hosted endpoint, sellingpartner-ai.amazon.com, over HTTP. There is nothing to install or host yourself.
- Seller Central login, not developer credentials. Authorization is the standard Seller Central OAuth consent flow. The assistant can only use tools your Seller Central role permits. Account owners manage who else may connect under a new Manage Agents page, found under Manage Your Permissions. Enabling the agent and authorizing secondary users are two separate switches, and both must be set.
- Eleven skills and three commands. The skills are instruction files that tell Claude which tools to call and in what order. The commands are shortcuts: a stockout check, a listing fix, and a sales-drop diagnosis.
- Preview, then approve. Amazon's own rules for a price change are a useful window into the write model: call the listing patch in validation-preview mode first, show the current and proposed value, use exactly the number the seller typed, wait for the word "approve", one change per approval, and no unsolicited follow-up. An accepted patch means submitted, not live, and the skill tells Claude to say so.
- Data handling, in Amazon's words. The skills send tool calls and their parameters to Amazon's endpoint, get your business data back, transmit nothing to non-Amazon endpoints, and do not persist your account identifiers or data beyond the session. Your prompt text itself is handled by Claude under Anthropic's terms.
What it does today
Working from the skill definitions, here is the real capability list as of early October 2026.
- Seller analytics. Read-only questions about FBA inventory health, traffic and sales: available quantity, days of supply, demand forecasts at 7, 30 and 60 days, restock date and quantity, inbound totals, sessions and page views by device, featured-offer share, ordered units and ordered product sales. The skill discovers which metrics your marketplace exposes before querying, because not every metric is available to every seller.
- Stockout prevention. Daily velocity from aggregate order metrics, days of cover from fulfillable quantity, a reorder quantity and date, and the inbound gap. Amazon's documentation notes there is no reports tool on the server, so velocity comes from the synchronous order-metrics call rather than the Sales and Traffic report.
- FBA inbound shipments, end to end. Create a plan, set packing, choose placement, choose transportation, confirm, with approval gates at each decision. It also handles labels, prep settings, delivery windows and box-content updates, and documents the constraints that trip sellers up, such as placement being permanent once confirmed.
- Listing troubleshooting. A router skill decides whether a listing problem is a reported issue, a buyability problem, or a searchability problem, and hands off to the matching specialist. Fixes are drafted in preview and need your approval.
- Compliance pre-flight. An advisory gate that runs before any compliance-sensitive listing write. It checks claims, ingredients, category, condition, images and identifiers and never writes anything itself.
- Team access and support. One skill walks an admin through granting agent access to secondary users; another prepares a clean, redacted summary for Seller Support and never submits it without explicit confirmation.
- Developer documentation. A read-only search over the SP-API docs, for people building their own integrations.
What it does not do, by design
These are not gaps we found by poking at it. They are exclusions written into Amazon's own skill files and launch coverage.
- No advertising. Every listing skill says "do not use for advertising". Ads live on a separate server, covered below.
- No financial reports. The analytics skill excludes tax and financial settlement reporting. Revenue here means ordered product sales, before refunds, before every fee Amazon charges, and before your cost of goods. Nothing in the plugin knows what a unit costs you or what you kept after Amazon's cut.
- No pricing advice. The stockout skill is explicit that it never suggests raising or lowering a price, never offers a price change as an option, and comments on nothing about whether a price is wise. If you name a number, it drafts exactly that number.
- One account, one marketplace per session. Every analytics call carries a single marketplace and a single seller entity. Multi-marketplace questions mean separate sessions or manual stitching.
- Third-party sellers only. The analytics domains carry a "Threep" prefix, Amazon's shorthand for third-party, and Vendor Central is out of scope in the launch coverage.
- No memory of what you approved. The skills do not persist data beyond the session. Seller Assistant's new memory, where Amazon says it follows you into Claude, is a feature of Amazon's assistant, not a decision log you can audit or query: there is no tool that lists what you approved last month and what it did.
- US only, Claude and Quick only, beta. Amazon says international expansion and more assistants will follow, without dates. Several names in the documentation, including the metric domains, are marked as subject to change before general availability.
One exclusion deserves its own paragraph because it is easy to miss. Amazon's documentation for the plugin notes that the server exposes no reports tool, which is why velocity comes from the order-metrics call instead of the Sales and Traffic report. A great deal of what sellers rely on only exists as a report: Search Query Performance, the Brand Analytics family (top search terms, market basket, repeat purchase, search catalog performance), settlement reports, the FBA inventory ledger, returns, reimbursements, storage and long-term storage fees, and the seller performance report behind account health. None of those appear in the published skills either. Customer review topics, seller feedback, A+ document bodies, Buy Box offers and best-seller rank are likewise absent. The Ads server, for its part, is advertising only. The matrix below is our reading of the three connectors as published; the half marks are explained in its footer.
The other half: the Amazon Ads MCP server
Amazon's advertising organization shipped its own MCP server into open beta on February 2, 2026, and it is still in open beta as we write this. It is a different product from a different part of Amazon, and the two do not share a data model.
- Who can connect. "Amazon Ads partners with active API credentials", in Amazon's words. In practice that means an Amazon Developer account, a Login with Amazon security profile, and an approved Ads API access application. That is a real barrier for a seller who has never built against the API, and a non-issue for one whose PPC tool already holds credentials.
- What it does. Reports and performance queries, account settings, billing data, and campaign management including creating, updating and deleting campaigns. The announcement describes campaign creation as producing a ready-to-launch campaign that only needs review, which is a workflow convention rather than an enforced approval gate.
- Where it runs. Hosted by Amazon with separate endpoints for North America, Europe and the Far East. Amazon names Claude, ChatGPT, Gemini, Kiro and Bedrock as clients.
- What it costs. No fee is published, and it is free during the beta.
- What it lacks. Seller data. No orders, no inventory, no listings, no fees, no profit.
Using both together: what works, and what breaks
You can add both to Claude today. You will have two connectors in the sidebar, each excellent at its half. For any question that lives entirely on one side, this setup is hard to beat and costs nothing: "which of my listings has an error right now", "how many days of cover do I have on this SKU", "which campaigns spent the most yesterday".
The trouble starts with the questions sellers actually run a business on, because each needs both halves joined on one product, one date and one currency:
- Is this campaign profitable after fees and my real cost of goods? The Ads side has spend and attributed sales. The seller side has ordered revenue, but no fees, no refunds and no COGS. Neither has the number you want.
- What is the break-even ACoS on this product this month? That needs settled revenue, Amazon's actual fees and your unit cost. None of the three is on either server.
- Should I cut spend before the stockout, and by how much? Days of cover is on one side, daily ad spend and conversion on the other, and the join is yours to do in chat, by hand, every time you ask.
- Did the bid change I approved three weeks ago help or hurt? Nothing on either side remembers that you approved it, let alone reads the before and after.
Two smaller frictions are worth knowing. The two servers scope differently: the seller plugin wants one marketplace per session, while the Ads server works per advertising profile and region. And the two approval models differ: the seller plugin previews every write and waits for the word "approve", while the Ads server leaves review to the client's workflow.
Knowledge, not only data
There is a second difference that the matrix cannot show. Amazon's skills carry rules for their own tasks: a stockout threshold at fourteen days of supply, a compliance checklist, a fix playbook for listing issues. What they do not carry is a view on how to run the account. There is no posture for a product, no margin band, no bidding or campaign-structure doctrine, no harvest or negation logic, no decision triggers, and the stockout skill says in plain words that it will not advise on price. That is a deliberate choice by Amazon, and probably the right one for a first-party tool: Amazon selling you a bidding strategy would be a strange thing. It does mean the plugin is a very good pair of hands and not an operator.
Our connection ships the operator's knowledge with the data. Before an AI proposes any optimization through it, the server instructs it to read our account management guide, a written methodology covering operating principles and guardrails, the financial spine of margin layers and profit per day, the product posture model, the advertising framework for campaign structure and bidding, the optimization sequence, AI-shopping-era optimization and a decision-trigger reference. The same doctrine is published free as The Amazon Operator's Field Guide, twenty-seven chapters on economics and break-even, the advertising loops of bids, placements, harvesting and negation, listings, inventory and forecasting, reimbursements and the operating calendar, also on GitHub chapter by chapter. It is the guide we run our own brands on, and the AI workforce's seats are written against it. Amazon's connectors hand your AI numbers. Ours hands it numbers and a way of working.
Where a third-party connection still earns its place
Here is the honest framing from a company that sells one. Amazon just made access to your own data free, native and well made. We think that is the right outcome, and it is why we moved to a single flat price last week: the plumbing is Amazon's job now. What is left to pay for is the layer the two Amazon servers do not have, and cannot have while they belong to two organizations with two data models.
- One warehouse, both halves, one key. We pull the same Amazon APIs the plugins use, store advertising and seller data side by side, and join them on product, date and currency so the question above is a single tool call.
- The data neither Amazon server exposes today. Search Query Performance by query and ASIN. Top search terms, market basket, repeat purchase and search catalog performance from Brand Analytics. Settlement transactions, financial events, refunds, reimbursements, the FBA inventory ledger, storage and long-term storage fees. Customer review topics and trends, seller feedback and the account health report. A+ document bodies, Buy Box offers, best-seller rank, Store analytics, hourly ad performance from the Marketing Stream. In all, more than sixty seller datasets and thirty-five advertising datasets your AI can query directly, with the warehouse enforcing which account it may see.
- Settlement truth plus your costs. Our profit and loss is built from settlement data, every fee and refund as Amazon posted it, with your cost of goods added, so every product carries a break-even ACoS and bids are judged against margin rather than ACoS alone.
- A decision queue with memory. Budget caps, listing issues, reorders, missing costs, account health and more file into one ranked queue with the evidence behind each item. Every approved change gets a write receipt and, once its attribution window closes, a verdict: helped, hurt or neutral. Your Claude can read that queue and those verdicts.
- One connector, every client. Our server exposes 121 tools to Claude, ChatGPT and any MCP client, across every marketplace of an account, and works for advertising-only accounts too. Exactly one of those tools can change your Amazon account, and it restates what you are approving first.
- An AI workforce that files into the same queue, so a seat's proposal and a platform signal sit next to each other instead of in separate tools.
The queue above is what the connector's desk tools return to your AI: the same ranked items, the same evidence rows. Below it, every approved change waits out its attribution window and is read for a verdict, which is the memory neither Amazon server keeps.
What we do not claim: that Amazon's plugin is worse at its own job. For FBA inbound workflows and listing repair it is thorough, it is free, and it speaks Seller Central's native tools. Use it. Then ask the question neither Amazon server can answer.
A quick decision guide
- The plugin alone is probably enough if you sell in one US marketplace, do not advertise or manage ads elsewhere, and mostly need inventory, inbound and listing help.
- Add the Ads MCP server if you hold Ads API credentials and want campaign reads and writes in chat, with the understanding that it knows nothing about your inventory or profit.
- Add a third-party connection when you need profit after fees and costs, break-even bids, cross-marketplace views, a decision queue, or a record of what your approvals did. That is the job we built for, and the live demo at amz-vault.com/demo shows it on a fictitious brand with no signup.
Sources
- Amazon, Selling Partner agentic toolkit (the plugin's source: README, plugin manifest, MCP configuration, skills and references), github.com/amzn/selling-partner-agentic-toolkit, read October 3, 2026.
- Anthropic, claude-plugins-official pull request 6293, "Add amazon-selling-partner", merged September 23, 2026.
- GeekWire, "Amazon opens its seller tools to outside AI agents, starting with Anthropic's Claude", September 23, 2026.
- TNW, "Amazon Seller Assistant plugin brings seller data to Claude", September 23, 2026, including the Westmoreland and Burns quotes and the Quick Plus terms.
- Epinium, "Amazon Accelerate 2026: Seller Central in Claude", for the Professional-plan, single-marketplace and Vendor Central scope notes.
- Amazon Ads, "Introducing the Amazon Ads MCP Server" (open beta announcement), February 2, 2026.
Published October 3, 2026. Both Amazon servers are in beta and both have said more is coming; we will update this page and move the date above when the facts change.