Amazon Opens Seller Tools to Third-Party AI Agents

Amazon has spent the past year restricting outside AI agents from acting on its marketplace, most visibly by blocking Meta's Muse assistant from completing purchases on Amazon.com in September 2026. Days later, at Amazon Accelerate 2026 on September 23, the company did the opposite for a different kind of agent: it opened the seller side of its business, pricing, inventory, listings, to outside AI tools through a new plugin, starting in beta with Anthropic's Claude and Amazon's own Quick assistant. The contrast is the story. Amazon is not opposed to agents acting on its platform. It is choosing, deliberately, which agents get to act and under what terms.

AWS

Anthropic

Claude

The Seller Assistant plugin launches with Amazon's own Quick assistant and, in beta, Anthropic's Claude, the model that already powers Seller Assistant inside Seller Central.

~90%Sellers already using outside AIPer Amazon, at Accelerate 2026

~60 secondsTo connect the pluginNo code required, per Amazon

12 monthsFree Amazon Quick PlusFor every primary account holder, through Dec 31 2026

>90%Seller Assistant recommendations acceptedPer Anthropic field CTO Eric Burns

What actually launched

Seller Assistant is the AI system built into Seller Central that already answers seller questions and, with approval, adjusts prices and listings. What changed at Accelerate is where sellers can reach it from. A new Selling Partner plugin, described by Amazon as taking about 60 seconds to connect with no coding required, pulls a seller's listings, inventory levels, sales analytics and performance metrics into an outside agent, launching with Amazon Quick and in beta with Claude. Once connected, a seller can check stockout risk, compare a proposed price against the Featured Offer and competing offers, or update a listing, without opening Seller Central at all. Amazon VP of worldwide selling partner experience Mary Beth Westmoreland put the intent plainly: "Our vision was that they would never have to log into Seller Central."

Two other pieces shipped alongside the plugin. Seller Assistant now carries persistent memory of a seller's pricing patterns, inventory cycles and growth goals, and that memory follows the seller across Seller Central, Quick and Claude instead of resetting between sessions. And a new canvas feature generates a visual workspace from live seller data on the fly, for comparing pricing strategies or diagnosing an underperforming ad campaign, redrawing itself as the seller asks follow-up questions. Amazon also paired the launch with an incentive: every primary selling-account holder worldwide can claim a free 12-month Amazon Quick Plus subscription through December 31, 2026, plus two more free seats for co-workers. The plugin itself is in beta for Amazon's U.S. stores only, with international markets to follow.

An Amazon fulfillment center in Ocala, Florida. The image shows the physical infrastructure behind Amazon's marketplace; Seller Assistant and its new plugin operate entirely in software, on top of this kind of backend.

Why Claude, and not Muse

Amazon's own framing for this launch, that roughly 90% of its sellers already use some form of outside AI to run parts of their business, is also the argument for why blocking every third-party agent was never a realistic long-term position. Sellers were already routing around Seller Central with general-purpose assistants; Amazon's move formalizes that behavior for one vendor rather than pretending it was not happening.

But the choice of vendor is not neutral, and it is worth naming plainly rather than glossing over. Amazon has invested up to $8 billion in Anthropic in earlier rounds and, as of an April 2026 deal, committed up to $25 billion more tied to Trainium compute and infrastructure, on top of Claude already being the model that powers Seller Assistant on Amazon Bedrock. Claude gets sanctioned, documented access with a 60-second onboarding flow. Meta's Muse, by contrast, was blocked days earlier for accessing Amazon.com without disclosing itself and for reportedly capturing customer credentials, a conduct-based objection Amazon has stated explicitly under its 2026 agent access rules requiring automated systems to identify themselves. Both explanations can be true at once: Amazon has a real, stated policy against undisclosed scraping-style agents, and it also happens to have a large financial stake in the one agent it chose to let in first. Judging this launch only against Amazon's stated policy, or only against its investment in Anthropic, misses that the two are operating together.

Amazon sellers make dozens of decisions a day on pricing, inventory, and ads. Sellers accept its recommendations more than 90% of the time.
Eric Burns, Anthropic field CTO

The permissions question this creates

Handing an outside agent the ability to change a live price or listing is a materially different grant than letting it answer a question, and Amazon's own design reflects that. Sellers choose in advance which data categories the plugin can read, and Amazon says every action, not just write actions, requires the seller's approval before it executes, with a complete audit trail behind it. That is a meaningfully more cautious default than an agent that simply acts and reports back afterward.

How an agent reaches a seller's Amazon account

The plugin connects a third-party agent to Seller Central data, but every write action still passes through a seller-approval gate before it reaches Amazon's backend.

Seller

Asks a question or requests a change in plain language, inside Claude or Amazon Quick.

Claude or Amazon Quick

The agent the seller chose interprets the request and calls the Seller Assistant plugin.

Seller Assistant plugin

Pulls listings, inventory and sales data; drafts the price, listing or inventory change.

Seller approvalapproval gate

The seller reviews and approves the specific action before anything executes. Nothing ships without this step.

Amazon backend

Seller Central systems apply the change and log it to the audit trail.

Sellers also choose, in advance, which data categories the plugin can read. Read-only questions can bypass the approval gate; any write action, a price, a listing or inventory change, cannot.

It does not remove the harder questions, though. Once a plugin is connected, seller account data is processed on infrastructure outside Amazon's own systems, and sellers, not Amazon or Anthropic, are the ones responsible for what an approved action actually does once it is authorized. An approval taken quickly, across dozens of pricing decisions a day, is a weaker check than one taken deliberately, and the responsibility for that gap sits with the person clicking approve rather than with either company's infrastructure. None of this is unique to Amazon. It is the same trade every platform faces the moment it lets an outside model act with a user's authority: the permission model has to be legible enough that approving is actually a considered choice, not a reflex click that becomes indistinguishable from full autonomy in practice.

A platform opening its backend, on its own terms

Strip away the specific vendor and what Amazon did is structural: a major platform exposed a defined slice of its backend, inventory, pricing, listings, to outside AI agents through a sanctioned interface, rather than leaving third parties to reach the same data by scraping a dashboard or reverse-engineering an API. That is a small, seller-specific instance of a broader pattern already playing out in consumer commerce, where platforms are building deliberate rails for agents to act on a person's behalf rather than pretending agents will not show up at all. The distinction that will keep mattering is exactly the one this launch surfaces: an agent operating through a documented, permissioned, audited plugin is a different category of risk than one operating undisclosed, even when both are technically capable of the same actions.

It is also a reminder that agent access to a platform is, for now, something platforms grant selectively rather than something agents have a standing right to. A seller who wants Claude's help today gets it because Amazon built a specific integration for Claude, not because Claude can reach any platform a seller happens to use. That is the gap tools built to be model-agnostic are trying to close from the other direction, so that which agent handles a task is a choice made on capability rather than on which integrations a given platform has decided to ship. Metir AI takes that approach across the wider set of tools people already work in, rather than tying a workflow to one assistant a platform happened to pick first.

Sources:

Image credits

Hero and in-body figure: "Amazon Fulfillment Center, Ocala" by Michael Rivera (Wikimedia user Mjrmtg), licensed CC0 1.0 (public domain). Source: Wikimedia Commons. Reviewed before publication; shows an Amazon fulfillment center building in Ocala, Florida, not a Seller Central or Seller Assistant product screen, as the captions state.