Amazon's new Selling Partner plugin lets Claude run stockouts, suppressions and listing fixes for you. It is useful, free and the same tool for every seller. The advantage now sits in what it cannot see: the shopper-side recommendation, off-Amazon authority and who approves the pricing calls.
Amazon Just Put Seller Central Inside Claude. It Will Run Your Account. It Cannot Win Your Category.
By Stephen Honight, Founder of Lmo7
On 23 September 2026, at Amazon Accelerate in Seattle, Mary Beth Westmoreland, Amazon’s VP of Worldwide Selling Partner Experience, said this about the company’s plan for Seller Central: “Our vision was that they would never have to log into Seller Central.” That is not a hedge. That is Amazon’s own account of where it wants sellers to end up, and it is worth taking at face value.
The announcement had three parts. A Selling Partner plugin that puts Seller Central’s data and actions inside external AI tools. Seller Assistant workflows that run around the clock rather than waiting for a seller to open a dashboard. And persistent memory inside Seller Assistant, so it keeps context between sessions. The plugin launches in Amazon Quick and in beta inside Claude, with more integrations promised (Amazon News, 23 September 2026; GeekWire, Todd Bishop, 23 September 2026).
My recommendation, before I get into why this matters, is simple: connect it. Brands running Amazon US should switch this on, in recommend-only mode first, when it fits their operation. It is a useful piece of free infrastructure. What I want to spend the rest of this piece on is what it does not do, because that is where the actual work now sits.
The short version, if you take nothing else away: switch the plugin on when it reaches your marketplace, and keep it in recommend-only mode until a named person owns the approvals. Every seller gets the same tool, so the routine account work it automates, the stock checks, suppression fixes and listing errors, stops being a way to win. The advantage moves to the side the plugin cannot see: what Rufus, Alexa for Shopping and ChatGPT actually say when a shopper asks for the best product in your category. And write down who approves what before the first workflow runs, because Amazon’s own numbers say sellers accept its recommendations more than 90% of the time.
What the plugin is, mechanically
The Claude marketplace listing describes it plainly: “Connect your Amazon Seller Central account to Claude. Ask about sales, inventory, and listing performance and get recommendations grounded in your own data, not generic advice. Review and approve every action, without switching back to Seller Central.” It is made by Amazon, verified by Anthropic, added to the marketplace in September 2026, sign-in required, and currently in beta. The connector sits at sellingpartner-ai.amazon.com/mcp.
Connecting takes about 60 seconds and needs no code. Access runs through the standard Seller Central OAuth consent flow, and the assistant can only use tools that a seller’s existing Seller Central roles allow (Amazon News, 23 September 2026; Amazon’s selling-partner-agentic-toolkit README on GitHub). The beta covers Amazon US stores only. International expansion is promised with no date attached (Amazon News; GeekWire, 23 September 2026).
The bundled skills, per Amazon’s own repo, are seller-analytics, stockout-prevention, fba-inbound-management, listing-troubleshooter, listing-issues, listing-buyability, listing-searchability (described as “search-content optimization with approval”), listing-compliance (advisory only, it never writes), invite-secondary-users, support-case-helper and sp-api-knowledge. There are shortcut commands too: /sp-stockout-check, /sp-fix-listing, /sp-sales-drop. Write actions are always human in the loop, drafted for approval rather than executed automatically (amzn/selling-partner-agentic-toolkit README).
Amazon’s own framing of the workflows is useful because it shows the ambition. Its quoted example: “Monitor my top category for competitive openings. If you see one, adjust my pricing and refresh the listings.” Sellers choose whether a workflow only recommends or also acts, every action needs approval, and every action is audit-trailed (Amazon News, 23 September 2026). Amazon says 90% of selling partners already use third-party AI tools, that Seller Assistant has hundreds of thousands of active users, and that sellers accept its recommendations over 90% of the time. These are Amazon’s own figures, reported by Amazon, not independently audited, and I am flagging that every time they appear because the acceptance rate in particular is going to matter later in this piece.
Two context points sit behind all of this. Seller Assistant runs on Claude models via Amazon Bedrock, and Amazon is a major investor in Anthropic. Amazon also announced free 12-month Amazon Quick Plus for every primary account holder worldwide, available through 31 December 2026, plus access for two co-workers (Amazon News, 23 September 2026).
Why a shared agent flattens the ops layer
Here is the mechanism, and it is worth being precise about it because the commercial implication depends on it.
Every seller who connects the plugin is running the same model family against the same underlying data, through the same skill set, and largely accepting the same recommendations. That is not a small thing. It means the fix for a stockout, the fix for a suppressed listing, the fix for a buyability issue and even the “search-content optimisation” edit that listing-searchability proposes will look broadly similar across sellers who are exposed to the same problem at the same time. When the diagnosis is shared and the fix is shared and it is available to everyone at no marginal cost, doing it well does not make you faster than the seller next to you. It stops you falling behind them.
That is a real benefit. Account hygiene, meaning the unglamorous work of keeping stock flowing, listings live and errors fixed, used to take a person’s time, some SP-API knowledge and a certain amount of institutional memory about where the traps are. Amazon is now handing that out for free, through a model the seller may already trust from using it elsewhere. My sense is that is a gift worth taking without much hesitation.
It is not a strategy though. Table stakes and advantage are different things, and the plugin sits squarely on the table-stakes side of that line.
What the plugin cannot see
This is the part that actually matters for anyone running a serious Amazon business, so I want to go through it properly rather than gesture at it.
The plugin reads and acts on your account. It does not show you the other side of the conversation, which is how AI shopping surfaces answer a shopper who asks “best X for Y” in your category, which brand gets named, and why. The listing-searchability skill, per Amazon’s own description, optimises for search-content, meaning being found. That is a different job to being recommended, and we have written about the gap between the two before, both in how Rufus’s answers get redrawn overnight (https://www.lmo7.com/blog/rufus-nightly-redraw-70764-queries-2026) and in how to measure Rufus’s actual influence honestly rather than by proxy (https://www.lmo7.com/blog/measuring-amazon-rufus-honestly-2026). A listing can be perfectly indexed and still lose the recommendation. The plugin has no visibility into that outcome at all, because it is not a skill Amazon has published.
Everything off Amazon is also outside its reach. Reviews, third-party citations, retailer and publisher presence, your own D2C content. Lmo7’s view here has not moved: there is a quick-win track on the Amazon listing itself and a longer game on the authority that AI systems draw on when they build an answer, and this plugin only ever touches the first of those two tracks.
Ads and Brand Analytics, Amazon’s own reporting on what shoppers search and buy, are also absent from the published skill list, at least as of 24 September. Ad spend is where a meaningful part of Amazon’s own commercial incentive sits: GeekWire’s reporting of Amazon’s Q2 2026 results put quarterly advertising revenue at $19.8bn, up 26%. I am not going to draw a conclusion from that fact sitting next to the plugin’s scope. I will just put it in front of you and let you draw your own. We have written before about how Amazon Ads already sits inside the AI shopping conversation (https://www.lmo7.com/blog/amazon-ads-already-in-the-conversation-agentic-shopping-2026), and Alexa’s own move from voice toy to transaction surface follows the same logic of Amazon-owned surfaces sitting closest to the money (https://www.lmo7.com/blog/alexa-transaction-surface-not-voice-toy-2026).
And then there is the geography. This is US-only. For most Lmo7 clients reading this, the honest framing is that it is a preview of where the tooling is going, not something you can turn on this week. Use the gap to get roles and approval ownership sorted, so the switch-on is a formality when it lands.
The two-sided wall
In the same fortnight that Amazon opened its seller tools to Claude, it also blocked Meta’s Muse shopping agent from its store and said that outside agents must identify themselves and follow site rules (GeekWire, 23 September 2026). Put the two decisions side by side and the shape becomes clear. Amazon decides which agents are allowed to run a seller’s account, and it also decides which agents are allowed to shop its store. Operators get the agent Amazon chose for them. Shoppers get the agent Amazon owns, principally Alexa for Shopping. A brand selling on Amazon and direct-to-consumer is now dealing with an agent layer that Amazon controls at both ends of the transaction, from the seller side and from the shopper side. Neither side of this is neutral infrastructure, and any brand planning around agents needs to hold both decisions in view at once.
Governance is the new ops work
The risk here is not a rogue agent. Amazon has built approval gates into every write action, and listing-compliance is advisory only, never writing anything itself. The risk is that a 90% acceptance rate, Amazon’s own figure, turns the approval step into a rubber stamp. Amazon’s own workflow example, “adjust my pricing and refresh the listings” the moment a competitive opening appears, is a margin decision and a brand decision wearing the costume of an automation. Nobody should be approving that at the same speed they approve a stockout fix.
My recommendation would be a written agent policy, in place before the first workflow goes live rather than after. It needs to cover which action classes can run in recommend-only mode, which need a named human approver, who actually holds the Manage Agents roles in Seller Central, how a listing write gets checked against claims and compliance before it ships, and a weekly review of the audit trail measured against contribution margin rather than headline sales. That last point matters more than it sounds. A pricing change that lifts units sold can still be the wrong call if it was never checked against what it does to margin.
Where this leaves Lmo7’s own view
To be clear about where this piece is coming from: Lmo7 already runs Amazon account data through Claude over MCP as part of its own delivery work, so this is not commentary from the outside. We have lived with what this kind of surface shows and what it leaves out, and the gap described above is drawn from that, not from Amazon’s marketing.
The plugin is not something to be defensive about. It is a fair, useful piece of infrastructure, and being fair to it is part of what makes this argument credible. The honest position is this: the plugin takes the repetitive operational work off the table, and what remains is the work that decides who wins the category. Content that earns the AI recommendation rather than just the search index. Visibility measurement across Rufus, Alexa for Shopping and the wider AI surfaces. Hands-on optimisation and the judgement behind each approval. That is where the attention should go now, and it is work the plugin will not do for you. It is also work that moves the numbers: when Lmo7 ran that recommendation-side programme for Trip Drinks, the brand’s average position across the major LLMs went from 7th to 3rd and AI referral traffic rose 33%, measured across Lmo7’s own visibility tracking over the 60-day engagement.
What I would do next
Connect it, in recommend-only mode first. US sellers can do this now. UK and EU brands should use the wait to get roles and an approval owner sorted in Seller Central, so there is no scramble when it arrives. Owner: head of ecommerce. Why: free account hygiene and faster diagnosis, without handing over margin decisions by default.
Write the agent policy before the first workflow runs, not after. Set out the action classes, who approves each one, the compliance pre-check on any listing write, and a weekly audit-trail review measured against margin. Owner: ecommerce lead with finance. Why: Amazon’s reported 90% acceptance rate, combined with an automated pricing workflow, makes the approval step the likeliest point of failure.
Measure the side the plugin cannot touch. Baseline how Rufus, Alexa for Shopping and ChatGPT answer the buying prompts that matter in your category today, then put the content and authority work in place that changes those answers. Owner: brand or growth lead. Why: once account hygiene is the same for every seller, the recommendation is the only place left for share to move.
For brands set up to run this themselves, Lmo7’s DaaS tier (£250 + VAT per month per marketplace) gives direct API access to the account data plus the training to use it, alongside the visibility and measurement layer the plugin cannot see. For brands that want it run for them, the Challenger tier covers the recommendation side: Amazon Content Optimisation at £750 per month and Agentic Tracking & Content Optimisation at £1,000 per month, with the agent policy and approval discipline built in alongside. For multi-brand portfolios, the real question is who owns the agent roles and the approval chain across brands and markets, and that conversation starts with an Enterprise workshop, from £5,000.
Stephen Honight is the founder of Lmo7, an agency helping brands including Trip Drinks, Veloforte, Brown-Forman, Haleon, Pelotan and Symprove win visibility and recommendation share across Amazon and AI shopping surfaces.