Marketing teams already run Claude for briefs, ad copy, and data pulls. Amazon sellers have been slower. But that gap is closing fast. Sellers now connect Seller Central and Amazon Ads directly to Claude, then ask questions in plain English instead of exporting another spreadsheet and building another pivot table.
The connection layer is the Model Context Protocol, or MCP. It gives Claude a live read on your account data instead of a static file you uploaded last Tuesday.
Five ways to connect your Amazon seller account to Claude in 2026
You connect Amazon to Claude through a third-party MCP server, Amazon's own Ads MCP server, the Selling Partner API directly, a Google Sheet, or BigQuery.
Anthropic has no first-party Amazon connector, and Amazon's official MCP covers advertising only. Everything else runs through a bridge you either buy or build. Which bridge depends on how much data you move and who maintains it.
| Method | Setup effort | Data freshness | Best for | Watch out for |
| Third-party MCP server | 15 minutes, no code | Live to daily refresh | Sellers who want answers this week | Monthly fee; the vendor holds your OAuth token |
| Amazon Ads MCP server | Ads API credentials required | Live | Partners and agencies already on the Ads API | Advertising data only, no Seller Central |
| Direct SP-API build | Days of engineering work | Live | Teams shipping Amazon data as a product | You own OAuth, throttling and pagination |
| Google Sheets or CSV | 10 minutes | Daily schedule or manual | Finance review before Claude sees numbers | Claude aggregates in-sheet; totals can drift |
| BigQuery | Warehouse and ETL setup | Scheduled loads | 10+ accounts or years of history | Needs SQL and someone who owns the pipeline |
Method 1: Connect Seller Central to Claude with a third-party MCP server
This is the fastest route and the one most sellers should start with. No code, no developer approval.
Vendors including DataDoe, Windsor.ai, Seller Labs, Two Minute Reports and Adzviser host MCP servers that sit between Amazon's Selling Partner API and Claude. You authorize once through Amazon's OAuth flow and they handle the rest.
Pricing sits where most mid-tier Amazon Ads Software does. DataDoe, as one example, is $97 a month with a 14-day trial and every integration switched on.
Step 1: Authorize Amazon data access
Your MCP provider needs permission on Amazon's side before Claude can read a single row of data.
- Pick a provider that hosts an Amazon MCP server. DataDoe, Windsor.ai, Seller Labs, Two Minute Reports and Adzviser all do.
- Inside the provider dashboard, add Amazon Seller Central or Amazon Ads as a data source.
- Complete the Amazon Selling Partner OAuth flow and approve the read scopes. Approve write scopes only if you want Claude changing bids and listings.
- Wait for the first sync. Most providers backfill 30 to 60 days before the connector reports as ready.
Step 2: Copy your MCP connection details
The provider generates a remote endpoint tied to your authorized account. Treat that URL like a password, because anyone holding it can read your account.
- For Claude on the web or desktop, copy the MCP server URL from the connector dashboard.
- For Claude Code, copy both the MCP key and the server URL. The key usually goes in an Authorization header.
- Note which marketplaces the token covers. A US-only authorization will silently return nothing for your UK ASINs.
Step 3: Add the connector in Claude
Custom connectors work on Free, Pro, Max, Team and Enterprise plans. Free accounts get exactly one.
On Pro or Max, open Settings, go to Customize, then Connectors. Click the plus icon, choose Add custom connector, paste your server URL, and add OAuth credentials under Advanced settings if your provider issued them.
Team and Enterprise work differently. An owner adds the connector under Organization settings first, then each member connects it individually from their own Connectors page.
Running the desktop app with a local server instead? Then edit the configuration file on disk directly.
- macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
- Windows: %APPDATA%\Claude\claude_desktop_config.json
Paste your provider's configuration object inside the mcpServers block, save the file, and restart the app. The connector will not appear until that restart finishes.
Step 4: Add the same server to Claude Code
Skip this step unless you write scripts against your account. It is optional for everyone else.
claude mcp add --transport http amazon-seller \
https://mcp.yourprovider.com/mcp \
--header "Authorization: Bearer YOUR_MCP_KEY"
Type /mcp inside Claude Code to confirm the server connected. A server listed with a url field but no type field gets skipped silently.
Step 5: Run your first queries
Test with questions you already know the answer to. That is how you catch a broken mapping early.
- "What were my total units ordered and total sales in the last 7 days?"
- "List my top 5 ASINs by units ordered this month."
- "How many sellable units of SKU ABC-123 are in FBA right now?"
If the numbers match Seller Central, the connection is sound. If they are close but not identical, check whether the provider is reporting on order date or ship date.
Method 2: Use Amazon's official Ads MCP server for advertising data
Amazon shipped its own MCP server for advertising, and it went to open beta on 2 February 2026.
It covers campaign creation, updates and deletion, performance reporting, account settings and billing. Pre-built tools handle multi-step work such as generating a report, launching a campaign, or expanding into a new marketplace.
Access is limited to Amazon Ads partners holding active API credentials, available globally. If you have never touched the Ads API, this route is closed to you today.
The hard limit is scope. This server does not read Seller Central, so inventory, orders, returns and fee data stay outside it. You get everything the Amazon Ads Dashboard shows and nothing from the operations side.
It does handle creative-heavy formats well, including Amazon Video Ads reporting that most third-party connectors flatten or skip.
Method 3: Build directly on the Selling Partner API
Going direct means registering as an Amazon developer, running Login with Amazon OAuth yourself, and calling SP-API endpoints from your own code or from Claude Code.
The payoff is control. Every endpoint, every parameter, no vendor deciding which fields matter. The cost is that you now own refresh tokens, pagination, schema changes, retry logic, and the throttling behaviour underneath all of it.
Throttling is the part engineering teams underestimate. SP-API runs a separate token bucket for every operation.
A rate limit of one request per second with a burst of two means you fire two calls immediately, then wait. Empty bucket returns a 429.
Some plans are dynamic, adjusted per seller against business metrics rather than how often you call. Read the x-amzn-RateLimit-Limit response header where it appears, and build your retry backoff before you build any features.
This route suits engineering teams shipping Amazon data as a product feature. It suits almost nobody else.
Method 4: Route reports through Google Sheets or a CSV
Push Amazon data into a single Google Sheet on a daily schedule, or export a report by hand, then point Claude at the Sheet.
The trade-off is real and worth stating plainly. With MCP, Amazon does the filtering and aggregation server-side and hands Claude a clean result. With a Sheet, Claude aggregates across the rows itself.
At a few hundred rows that is fine. At forty thousand, totals and averages start drifting.
Speed is the upside. A pre-built Sheet covering two years of order history returns faster than any live API pull, and finance can review the numbers before Claude sees them.
One reporting quirk to plan around: the Sponsored Products search term report only includes terms that received a click inside your selected date range. Terms with impressions and no clicks never appear.
The lookback window also varies by report and account, so pull weekly and keep your own archive.
Method 5: Load into BigQuery when you outgrow live pulls
Once you run several large accounts, live API pulls stop working well. Claude will time out or tell you the request is taking too long.
BigQuery fixes the shape of the problem rather than the symptom. You load Amazon data on a schedule, pre-aggregate it into small summary tables, then point Claude at those tables through a BigQuery connector or SQL in Claude Code.
Claude then queries a few thousand summarized rows instead of pulling millions through a rate-limited API.
This is the route for agencies running ten or more accounts, and for Amazon Ads for Scaling Brands where a year of hourly data is normal.
Give Claude the context Amazon does not store
Amazon knows your sales. It does not know your cost of goods, and neither does Claude.
Set up a dedicated Claude Project and load it with the numbers that decide every recommendation: landed COGS per SKU, supplier lead times, target contribution margin, brand voice rules, and the competitor ASINs you actually track.
Without that file, Claude will hand you an ACoS number and call it profitability. With it, you get net margin after fees, freight and ad spend.
Add a note on which campaigns are defensive. Otherwise Claude will recommend cutting the ones protecting your branded terms.
Check Amazon's agent policy before you connect anything
Amazon updated its Business Solutions Agreement effective 4 March 2026, and that update addresses automated agents directly for the first time in its history.
Three baseline obligations apply. Agents must identify themselves as automated systems at all times, follow the new Agent Policy, and cease access immediately if Amazon asks. Amazon can also restrict access at its own discretion.
The definition of an agent is still loose, and that vagueness is where sellers keep getting stuck.
Sellers on the forums have asked whether order-polling software and FBA reimbursement tools now count. Amazon has published no threshold and no clear answer.
Two practical rules hold up regardless. Pick providers running on official SP-API access rather than anything scraping Seller Central, and keep every write action behind a human approval step you can point to in an audit.
Six things Amazon sellers should hand to Claude first
Start with the work that is repetitive, data-heavy, and currently eating a full afternoon every week.
1. Pull negative keywords out of the search term report
The search term report is where most wasted ad spend hides, and filtering thousands of rows by hand is why nobody does it weekly.
"Find search terms from the last 30 days with more than 15 clicks and zero sales, or an ACoS above 65%. Group them by campaign and format the output as a bulk upload sheet for negative exact keywords."
Claude returns the terms and formats them for bulk upload, so the whole loop runs in minutes rather than an afternoon in the Advertising Console. Pair it with a monthly review of your Amazon PPC Ads structure.
2. Flag stockout risk against real sales velocity
Stockouts cost you organic rank, and losing rank costs far more than the missed sales did.
Claude compares your current FBA sellable quantity against sales velocity over rolling 7, 14 and 30-day windows, then reads that against your supplier lead time.
"Compare my current FBA sellable quantity against 14-day sales velocity. Flag any ASIN that hits zero inside my 30-day supplier lead time, and give me the reorder quantity for 60 days of cover."
The output you want is a reorder quantity and a date, not a warning. Velocity alone will mislead you badly during a Prime event, so tell Claude which promotional windows to exclude.
3. Mine competitor reviews for listing angles
Paste a competitor's one and two-star reviews into Claude alongside your own product details, then ask what specific flaws buyers keep complaining about most.
"Here are 40 one and two-star reviews for a competitor ASIN. Identify the three most repeated product complaints, then rewrite my bullet points to answer each one directly without naming the competitor."
Three recurring complaints about a rival's sizing is a bullet point you should be writing this week.
This works because review text is unstructured, and unstructured text is where Claude beats a reporting dashboard outright. No PPC tool will ever tell you that customers hate a competitor's packaging design.
4. Rewrite listings against terms you actually rank for
Claude cross-references your current title and backend search terms against the terms your own ads already convert on, then rewrites the copy around them.
"Review the title and backend search terms for ASIN B0XXXXXXX. Cross-reference against my top 20 converting search terms from PPC. Rewrite the title to include the three highest-volume terms in natural sentence order, under 200 characters."
The value is in the pairing. Volume data alone produces keyword-stuffed titles, while your PPC conversion data tells you which terms buyers act on. Read our guide to Amazon Listing Keyword Optimization first.
Rufus is the other reason this matters now. Amazon's AI shopping assistant reads listing copy as answers to shopper questions, not as a keyword field.
5. Turn return reasons into fewer returns
Return reasons are free product feedback that nobody ever reads. Claude will read every single one.
Feed it 30 days of return reasons and negative seller feedback, then ask what the single most common point of confusion actually is.
"Pull return reasons and negative feedback from the last 30 days. Rank the causes by volume, then draft an insert card and an A+ content block that address the top cause before the customer buys."
A return rate that drops from 8% to 5% shows up in your Account Health Rating and your margin at the same time. Insert cards, image updates and clearer sizing charts are the usual fixes.
6. Audit true net margin after Amazon fees
FBA fees, referral fees, storage fees and ad spend all move constantly, and gross revenue tells you nothing about which SKUs are actually earning money.
"For ASIN B0XXXXXXX last quarter, take gross sales and subtract referral fees, FBA fulfilment fees, storage fees, ad spend and $4.50 COGS per unit. Return net margin percent and flag if ad spend exceeds 20% of revenue."
Claude does the subtraction across every fee line and returns net margin per ASIN. Cross-check the ad spend figure against what you expect to pay, since Amazon Ads Cost varies enough by category to distort the result.
Run it monthly. Products slip quietly from profitable to break-even after a single Amazon fee change.
What Amazon data you can query once the connector is live
| Group | What you can ask about |
| Orders | Amazon Order ID, order and ship dates, order status, currency, earliest delivery and ship date, invoice status |
| Items | ASIN, SKU, condition, quantity ordered and shipped, item price, promotion discounts, COD fees |
| Sales | Total sales, order count, order item count, unit count, average unit price |
| FBA inventory | Sellable quantity, reserved, inbound, customer damaged, carrier damaged, defective, expired, fulfilment network SKU |
| Financials | Referral fees, FBA fulfilment fees, storage fees, refunds, reserves, reimbursements, settlement events |
| Listings | Listing status, created date, main image URL, title, bullet points, backend search terms |
| Buyers and shipping | Buyer name and email, purchase order number, shipping city, state, country, postal code |
| Advertising | Campaign, ad group, keyword, search term, impressions, clicks, spend, orders, sales, ACoS, ROAS |
Ad data comes through the same connection on most platforms, which is what makes blended questions possible. Querying Amazon Sponsored Ads beside organic order data in a single prompt is the whole point.
Breakdowns usually include date, fulfilment channel, marketplace and order status, plus campaign type for Amazon Sponsored Display Ads.
Where Claude stops and an operator takes over
Claude is good at reading data and bad at knowing when the data is wrong. That distinction decides how much you should trust it.
It will average a column containing a blank month without flinching. It will treat a cancelled order as revenue if the table says so. It will not question an unusual fee line, because it has no memory of last quarter.
So verify totals against Seller Central before you act on them. Every time, for the first month.
Write actions need the same discipline. Bid changes, listing edits and repricing should sit behind approval gates, both for your own safety and for Amazon's agent policy.
The honest limit is that a chat window is not an operating system. Claude answers the question you thought to ask, on the day you thought to ask it, which is a different thing from an account being managed hourly.
That gap is why sellers running real volume pair this setup with Amazon Ads Management Services rather than treating a connector as the whole answer.
Comparing your options? Our list of Amazon Product Ads Management Companies covers who actually does this well.
Running this yourself in a chat window works. Xneeti runs the same analysis hourly with a dedicated strategist reviewing every AI decision before it touches your account. Book a demo to compare.
