Amazon opens Seller Central to AI agents: Claude and Quick bots will manage your e-commerce store
🔎 The sell-side enters the agent era
On September 23, 2026, on stage at Accelerate, Amazon's annual seller conference in Seattle, the company did what nobody expected from the most closed-off e-commerce platform on the market: open Seller Central to third-party AI agents. A beta plugin, initially reserved for US sellers, makes it possible to manage inventory, pricing, listings, and analytics directly from Claude or from Quick, Amazon's in-house assistant — without ever opening the web interface.
"Our vision was that they would never have to log into Seller Central," summarizes Mary Beth Westmoreland, VP Worldwide Selling Partner Experience, quoted by GeekWire. The statement sounds innocuous. It sounds like a death sentence for manual data entry.
Why now? Because the practice already existed, under the radar. About 90% of Amazon sellers already use third-party AI, often through unofficial bots that scrape the interface. Amazon isn't creating a behavior: it's recognizing it, framing it, and reclaiming it. AI Weekly on September 24 talks about a shift, not just a product update — that's exactly the right framing.
The financial stakes explain everything else. Fees collected from sellers brought in $46.8 billion in Q2 2026 alone, more than AWS's revenue (GeekWire figure). A surface that large couldn't be left as a poaching ground for rogue agents.
The Essentials
- A beta plugin (US) connects Claude and Quick to the Seller Central APIs: inventory, pricing, listings, sales analytics. Connection in ~60 seconds, without a single line of code.
- The enhanced Seller Assistant runs on Amazon Bedrock, combining Amazon's Nova models and Anthropic's Claude, with nearly 25 years of marketplace data. At no additional cost for all US sellers.
- 12 months of Quick Plus free for any primary seller account holder, offer valid until December 31, 2026.
- Three-layer governance: permissions by data type, human approval for every action, complete audit log.
- ~90% of sellers were already using third-party AI: Amazon is formalizing a widespread practice.
- Tense backdrop: Amazon blocked Meta's Muse shopping agent a few days before the announcement; external agents must now identify themselves and follow the site's rules.
Recommended Tools
| Tool | Main use | Price (month year) | Ideal for |
|---|---|---|---|
| Claude | Manage inventory, pricing, listings and analytics via the Amazon plugin | Depends on subscription (September 2026, check on anthropic.com) | Sellers who want a third-party agent with a full audit trail |
| Quick / Quick Plus (Amazon) | Native assistant, background automations, persistent memory | 12 months free for any primary account before 12/31/2026 | New sellers starting out with automation |
| Seller Assistant (Amazon Bedrock) | Inventory optimization, compliance, advertising campaigns | At no additional cost (September 2026) | All US sellers, no exceptions |
| Hostinger | Build an independent e-commerce site to diversify away from the marketplace | From ~$3/month (September 2026, check on hostinger.com) | Sellers who want to reduce their dependence on Amazon |
What Amazon actually announced (and what it didn't announce)
Amazon has launched a beta plugin for US sellers, connecting Claude and Quick to the Selling Partner APIs. Not an open protocol, not a public API: a controlled door, inside its ecosystem, on its terms.
Announced on September 23, 2026 at Accelerate (Seattle), the plugin covers the seller's day-to-day: inventory management, price adjustments, creating and editing listings, and reading sales analytics (SiliconAngle). Connecting to Claude takes about 60 seconds, no code required, and the seller chooses precisely which data the agent can view and which actions it can perform (GeekWire).
What it isn't: an open standard. Amazon deliberately ruled out the protocol route. Shopping-agent standards — UCP, ACP, AP2 — remain confined to the buy-side, and no one has imposed a format on the sell-side (Ideabosque). The architecture is described as "modular," with further integrations promised in the coming weeks.
Claude isn't a neutral choice either. Amazon is a major investor in Anthropic, and Claude already powers the internal Seller Assistant. The plugin therefore opens a door… to a long-standing partner. The openness has contours, and they are capital-driven.
The three surfaces open to sellers
| Surface | Who it's for | What it does | Cost |
|---|---|---|---|
| Claude / Quick plugin (beta) | US sellers | Inventory, pricing, listings, analytics from the agent, without opening Seller Central | Free (Claude subscription required) |
| Seller Assistant (Bedrock) | All US sellers | Inventory optimization, compliance, ad campaigns, persistent memory | At no additional cost |
| Quick Plus | Primary seller account holders | Background automations, enhanced assistant | 12 months free through December 31, 2026 |
Why this is a turning point for agentic commerce
Because it's the first time a marketplace of this size has officially opened its "seller" side to autonomous agents. Agentic commerce is no longer limited to agents that buy: from now on, there will be agents that sell too — supervised and identified.
Until now, the agentic wave had been focused on the buy-side. The UCP, ACP and AP2 protocols organize the agents that buy on behalf of their user. General-purpose assistants are learning to work alone for hours — like ChatGPT Work, OpenAI's agent that works for hours without you. But the sell-side remained a protocol wasteland (Ideabosque).
Amazon has just colonized it, in its own way: no open protocol, but a controlled surface, granular permissions, and an audit log. The message to third-party sellers — those who account for more than 60% of the platform's sales — is crystal clear: your bots are now legitimate, provided they come through the front door.
My take: the plugin itself is almost secondary. The signal is the legitimization. When the world's number one in e-commerce declares agents welcome as long as they identify themselves, bots that scrape the interface become a compliance problem overnight. Seller tool vendors no longer have a choice — align or disappear.
The movement is only just beginning. Autonomous agents are becoming widespread, as shown by our selection of the best autonomous AI agents, but hooking them into systems that touch real money remained the exception. Amazon has just normalized the exception.
Seller Assistant: the in-house agent that has read 25 years of marketplace
Seller Assistant is Amazon's native agent: it runs on Amazon Bedrock, combining the company's own Nova models with Anthropic's Claude. It's the one Amazon pushes by default, and it's available immediately to all US sellers.
Its playground: inventory optimization, compliance monitoring, and advertising campaign management, backed by nearly 25 years of marketplace data (Crypto Briefing). Its execution remains subject to seller authorization — Amazon repeats this in every sentence, a sign that the topic is sensitive both internally and externally.
The numbers give a sense of the ambition. Previous generic AI tools reached about 1.3 million sellers. Today, sellers accept ~90% of AI recommendations. And the Canvas Experience, an interactive visual dashboard launched in March 2026, had already gotten sellers used to steering their business outside the classic interface.
The real novelty is persistent memory. Seller Assistant remembers your pricing patterns, your inventory cycles, and your growth objectives (GeekWire). It no longer just answers: it anticipates, drawing on what it has learned about your habits between sessions.
This is exactly the direction the entire agentic ecosystem is taking. Our article on Anthropic Dreaming, where Claude agents learn from their "dreams" between sessions, describes the same trajectory: memory is no longer a bonus, it's the product. An agent without memory remains an eternal intern; an agent with memory becomes a collaborator.
On the model side, the choice of a Nova + Claude mix on Bedrock is telling: combining an economical in-house model with a cutting-edge third-party model is becoming the norm for agentic AI, as detailed in our guide on the best LLMs for AI agents.
A word of caution, however, about that 90% acceptance figure: accepting a recommendation is not the same as validating its relevance. It's an indicator of trust, not of performance. Stay clear-eyed.
Permissions, approval, audit: the governance model dissected
Amazon imposes three layers of control: permissions by data type, human approval for every action, and a complete audit log. Understated, but it's probably the model that will prevail for any agent that touches money.
The seller defines the scope: the agent can read analytics but not touch prices, or the reverse. Every write action goes through explicit approval. And everything is logged, action by action, in a searchable audit log (Ideabosque).
It's a smart compromise between two extremes. Neither the total sovereignty of a bot that does whatever it wants, nor the black box that acts without leaving a trace: a limited, revocable, auditable mandate. Banks spent a decade imposing this level of rigor on fintech; Amazon imposes it on its agents from day one.
But technical guardrails aren't enough. A poorly instructed agent will execute a bad instruction flawlessly. The real protection is the context you give it: floor prices, minimum margins, listing tone, restocking thresholds. That's exactly the work described in Context files: CLAUDE.md, AGENTS.md and beyond — a well-written instruction file is worth a thousand approvals clicked in a row.
My advice: before connecting anyone to your seller account, put your business policy down in black and white. The agent won't invent your margins for you — and that's just as well.
The Meta Muse Affair: Openness Has a Boundary, and It's a Sharp One
Yes, Amazon blocked Muse, Meta's shopping agent, just days before the announcement. That's no coincidence: it's a double-edged message.
The contrast is telling. On the sell-side, Amazon opens wide: agents working for its sellers create value within its ecosystem and feed its commissions. On the buy-side, it shuts the door: shopping agents browsing en masse distort traffic, skew prices, and bypass advertising (GeekWire, SiliconAngle).
The rule being laid down is simple: external agents must identify themselves and follow the site's rules. Those that refuse get their access cut off. Muse found that out the hard way — proof that the rule applies to tech giants too.
Let's call things by their name: Amazon's brand of "agentic commerce" is not an internet of emancipated agents — it's a licensing regime. Open shopping protocols are advancing elsewhere, but on the marketplace, agent access will be conditional, traceable, and revocable at any time.
It's rational — nobody wants unidentified agents placing orders at millisecond speed across a catalog of several hundred million items. But don't confuse openness with neutrality: Amazon opens what it controls and locks down what it doesn't.
Margins, fraud, dependency: the announcement's blind spots
The announcement sidesteps three questions that will quickly become burning issues: margins, fraud, and liability.
Margins, first. If every seller plugs in an agent that optimizes their prices in real time, you get repricing wars between machines. Your agent adjusts against the competitor's agent, which adjusts against yours. The spiral goes down fast, and it's always the seller who absorbs the margin compression — never the platform, which collects its commissions on volume.
Fraud, next. An agent with write access to listings and prices is an attack surface. One compromised account, one malicious instruction slipped into the configuration, and a catalog can be sabotaged overnight. The audit log helps with investigation, but prevention remains the seller's responsibility.
Governance, finally. Who pays when an agent lists 500 SKUs at 1 euro? The seller approved the action, the agent executed it, the model misinterpreted the instruction. Legal and insurance frameworks don't exist yet. Amazon addresses this upstream with its permissions and audit trail, but final liability remains a blind spot.
Add to this a more insidious dependency: if all your operational memory — pricing patterns, inventory cycles — lives in Amazon's Seller Assistant, switching platforms will suddenly cost a lot more. Agent data portability is mentioned nowhere. That, in my view, is the real trap of the whole setup.
What this concretely changes for your business
Less time in Seller Central, faster pricing and stock reactions — and a new supervision workload to absorb. That's the honest assessment for a seller.
The plugin's immediate use cases:
- Repricing: the agent detects your pricing patterns and proposes adjustments that you approve — then let run in the background once trust is established.
- Inventory: anticipating restocking cycles based on your history, with alerts before stockouts.
- Listings: creation, optimization, and corrections without ever opening the web interface.
- Analytics: reading sales data and on-demand reports, in natural language.
- Compliance: continuous monitoring of marketplace rules via the Seller Assistant.
The economic context explains Amazon's haste. Third-party sellers generate more than 60% of the platform's sales, and their fees brought in $46.8B in Q2 2026 — more than AWS (GeekWire, Ideabosque). More efficient sellers mean more volume, more advertising, more commissions. Altruism has nothing to do with it.
For developers who want to build on these APIs, timing matters: agentic costs are taking shape right now, as shown by our analysis of Claude Code's move to monthly credits and what it means for devs and autonomous agents. Building business agents on these surfaces will be a profession — and probably a market.
One last, counterintuitive point: the more your store automates, the more your strategic independence matters. An agent manages your Amazon account, but it won't protect you from an arbitrary suspension, a commission hike, or competition from Amazon's own brands. Diversifying to your own site — for example hosted on Hostinger — remains the online seller's best life insurance.
How to Get Started: From Connection to Your First Automation
Three steps are all it takes: connect the plugin (60 seconds), restrict the accessible data, then approve the first actions before letting it run.
- Check your eligibility. The plugin is in beta for US sellers; international expansion is announced for "the coming weeks" (SiliconAngle), with no specific date for Europe.
- Connect Claude or Quick. Connecting takes about 60 seconds, without a single line of code (GeekWire).
- Scope the data. Start in read-only mode: analytics and inventory. Only open write access (pricing, listings) after a few days of observation.
- Approve each action. This is the default mode. Use it to fine-tune the agent's instructions: pricing rules, stock thresholds, product listing tone.
- Move on to background automations. Once trust is established, let repetitive tasks run and keep human approval for sensitive decisions.
- Activate the Quick Plus offer. 12 months free for any primary seller account holder, through December 31, 2026.
Budget-wise: the plugin isn't billed, Seller Assistant comes at no additional cost for US sellers, but access to Claude requires a subscription (pricing on anthropic.com, September 2026). Above all, plan for time: supervising an agent is work, not magic delegation.
❌ Common Mistakes
Mistake 1: Giving full access from day one
Connecting the agent with all permissions enabled is like handing your catalog to a 60-second-old intern. The solution: read-only for the first week, gradual write access, and pricing last of all.
Mistake 2: Confusing "90% acceptance" with "90% good decisions"
Sellers accept ~90% of AI recommendations (Crypto Briefing), but acceptance measures trust, not outcomes. Solution: sample the repricing decisions each week and compare your margins before/after.
Mistake 3: Ignoring the audit log
An audit log you never read protects you from nothing. Solution: weekly review of the agent's actions and clear documentation of what falls to humans versus machines — your accountant, and if applicable your insurer, will thank you.
Mistake 4: Putting all your sales on a single platform
Automation doesn't erase platform risk: suspension, rising commissions, competition from Amazon's own brands. Solution: a clean standalone site in parallel (Hostinger) and a multi-channel presence, with data you control.
❓ Frequently Asked Questions
Is the Amazon-Claude plugin available in France?
Not yet. The beta is limited to US sellers, with international expansion announced for the weeks following September 23, 2026 (SiliconAngle). No date for Europe has been communicated. Keep an eye on your Seller Central announcements.
How much does the whole thing cost?
The enhanced Seller Assistant comes at no additional cost for US sellers (Crypto Briefing). The plugin is not billed, but access to Claude requires a subscription (check anthropic.com). Quick Plus is free for 12 months for any main seller account created before December 31, 2026.
Is my seller data secure?
The permissions model limits access by data type, every action requires your approval, and everything is logged in an audit trail (Ideabosque). You choose what the agent sees and can revoke access at any time.
Why Claude and not ChatGPT?
Amazon is a major investor in Anthropic and Claude already powers the internal Seller Assistant (GeekWire). Since the architecture is modular, other agents will follow. To decide which one to use day to day, check out our comparison Claude vs ChatGPT.
Do I need to know how to code to connect my account?
No. Connecting to Claude takes about 60 seconds without a single line of code (GeekWire). Technical skills only become useful for advanced automations or custom integrations via the Selling Partner APIs.
Are shopping agents like Meta Muse blocked?
Yes. Amazon blocked Muse, Meta's shopping agent, a few days before the announcement (GeekWire, SiliconAngle). External agents must identify themselves and comply with the site's rules — the opening of the sell-side comes with no tolerance on the buying side.
✅ Conclusion
By opening Seller Central to AI agents, Amazon is turning an underground practice — 90% of sellers were already using third-party AI — into a regulated platform, and signing the birth certificate of agentic commerce on the seller side. The question is no longer whether an agent will manage your store, but which one: start with our selection of the best autonomous AI agents.
