Amazon has spent the past year watching its sellers route around it. By the company’s own estimate, roughly 90 percent of the merchants who sell on Amazon already use some outside AI tool to run part of their business, whether that means drafting listings, watching competitor prices or forecasting inventory. Rather than keep competing with that habit, Amazon decided at its Accelerate conference in Seattle on September 23 to build the on-ramp itself.
The company opened its Seller Central APIs to outside AI agents for the first time, launching a US beta plugin that lets sellers manage inventory, prices, listings and analytics through Anthropic’s Claude or Amazon’s own Quick assistant, without opening the Seller Central dashboard at all. It is, by a wide margin, the most access Amazon has ever granted a third-party AI system into the mechanics of running a store on its platform.
The short version
- Announced September 23 at Amazon Accelerate, opening Seller Central’s APIs to outside AI agents for the first time
- The new plugin works with Anthropic’s Claude and Amazon’s own Quick assistant, pulling listings, inventory and analytics into the conversation
- Seller Assistant runs on Amazon Bedrock, combining Amazon Nova and Claude models with persistent memory of a seller’s pricing and inventory patterns
- Amazon says connecting Claude takes about 60 seconds with no coding, though the beta is US-only, requires a Professional selling plan, and some sellers are waitlisted
- Every primary account holder gets a free 12-month Quick Plus subscription through December 31, 2026
- Sellers choose what data the plugin can reach and must approve every action before it executes
- Amazon left one thing untouched: the consumer-facing storefront, where buyers shop, stays closed to outside agents entirely
What sellers can actually do with it
The plugin’s job is to collapse a dashboard into a conversation. Once connected, a seller can ask Claude or Quick to check current stock levels, adjust a price across a product line, update a listing’s copy or images, or pull sales analytics for a given week, and the assistant acts on the account directly rather than just describing what to do. Amazon says the setup itself takes about 60 seconds and requires no code, a deliberate contrast with the Selling Partner API that developers have used for years to build custom seller tools, which typically takes real engineering time to integrate.
Underneath the conversational layer sits Amazon’s own Seller Assistant, which the company has quietly been building out well before this announcement. It runs on Amazon Bedrock, combining Amazon’s in-house Nova models with Claude, and the new workflows Amazon is rolling out let it monitor a business around the clock and flag or act automatically when predefined conditions occur, a stockout risk or a price falling outside a set range, for instance. Amazon says the assistant keeps a persistent memory of a seller’s specific pricing patterns, inventory cycles and stated growth goals across sessions, and that the underlying data never leaves Amazon’s own infrastructure even when the conversation happens inside Claude.
The permission model, and why it matters more than the plugin
What actually makes this launch notable is not that an AI can now touch inventory data. Plenty of third-party tools already do that through the existing Selling Partner API. It is that Amazon built a structured, public way for an outside model to request access and then act only within limits the seller sets. Sellers choose which categories of data the plugin can reach, and every write action, a price change, a listing update, requires explicit approval before it goes live rather than executing silently in the background.
That approval-gated design puts Amazon in the same camp as other large platforms that have spent 2026 building formal permission layers for agentic access rather than leaving it to each developer to improvise. Google took a similar approach when it opened its smart home platform to outside AI agents earlier this year, building an explicit connector and approval system rather than an open API anyone could call unsupervised. The pattern across both moves is the same: platforms are increasingly willing to let outside agents act on real accounts, as long as the platform keeps the approval switch.
The door Amazon left closed
The more revealing part of the announcement is what Amazon did not open. Industry analysts tracking the API surface, including a detailed breakdown from API Evangelist, pointed out that Amazon built genuine, well-documented agent access everywhere it owns the relationship and collects the revenue, inventory, pricing, listings, analytics, but drew a firm line at the consumer storefront itself. Where an agent would stand between Amazon and the shopper, the store stays human-only. No outside AI agent, not Claude, not any competitor, can browse products or complete a purchase on a customer’s behalf through this system.
That distinction says as much about Amazon’s priorities as the plugin itself does. Letting agents help sellers move faster costs Amazon little and likely increases the volume of listings and inventory turnover on the platform, which is good for Amazon’s marketplace revenue either way. Letting agents shop on a customer’s behalf, by contrast, would hand a layer of the buying decision, and potentially the brand relationship, to whichever AI company built the best shopping agent. Amazon, unsurprisingly, is not ready to give that up. The comparison is a useful one against Meta’s own agent push this week, where Muse is being built specifically to shop and check prices on a user’s behalf, the exact function Amazon’s storefront is staying closed to.
| Access point | Open to outside agents? | Notes |
|---|---|---|
| Inventory & stock levels | Yes | Read and adjust through the plugin, approval required for changes. |
| Pricing | Yes | Agents can propose and, with approval, apply price changes. |
| Listings & product copy | Yes | Editable through conversation, no Seller Central login needed. |
| Sales analytics | Yes | Full read access to performance and sales metrics. |
| Consumer storefront | No | Browsing and checkout remain human-only, no agent access. |
Based on Amazon’s September 23, 2026 announcement and independent API analysis. Beta scope may expand.
Signals to watch
- Whether the beta expands past the US and past Professional-plan sellers once the waitlist clears
- Whether Amazon adds more assistants beyond Claude and Quick, or keeps the integration list deliberately short
- Whether the storefront stays closed permanently, or whether competitive pressure from agentic shopping elsewhere eventually forces Amazon’s hand
- How sellers respond once the free Quick Plus year ends on December 31, 2026, and what the assistant costs after that
Amazon’s pitch to sellers is straightforward: stop tab-switching between a chatbot and a dashboard, and let the agent you already trust with market research also make the change. It is a genuinely useful shortcut for the roughly nine in ten sellers already leaning on outside AI. But the line Amazon drew between the back office and the storefront is the part worth remembering. The company is comfortable letting agents run the business behind the scenes. It is not yet ready to let one stand between Amazon and the customer.
