Thirteen Amazon MCP servers, ranked by criterion instead of overall. Nova wins profit modelling, SQP, marketplace coverage and multi-account. Amazon Ads MCP wins write actions. The open source repos win on free. Sources and dates on every claim.
Aug 18, 2026 · 14 min · 2,900 words
Thirteen servers an Amazon seller can connect this week. What each one reads, what it can change, and which one wins on the criterion you actually care about. Ecommerce platforms are covered further down; the money question here is Amazon.
The short answer
Test it in 5 minutes
You get rows: SKU, profit change, and the fee, price or ad-spend line that moved it.
Start the free trialOne overall ranking on a vendor's own page is worth nothing, so here is the ranking split by the thing you are buying for. Nova wins four rows. It loses two, and it loses them cleanly.
Deepest profit modelling
Net profit reconciled to the cent at SKU level across 40+ Amazon fee types, with your own COGS, FBM cost and VAT categories included.
Closest alternative: DataDoe models SKU profit but on daily tables and without your landed cost.
Search Query Performance
SQP exposed through the assistant, de-duplicated, with like-for-like period comparisons. Almost no other Amazon MCP exposes SQP at all.
Closest alternative: Helium 10 covers keyword and reverse-ASIN research, which is a different question.
Marketplace coverage and history
23 marketplaces in a single call, currency-normalised, with 24 months of backfilled history.
Closest alternative: Fluvi publishes broader marketplace coverage but not the history depth.
Agencies and multi-account
Several Seller Central accounts under one connection, one currency, one model, so a client question is one call rather than one call per account.
Closest alternative: Two Minute Reports puts Amazon beside other channels, but without the account economics.
Campaign write actions available today
Amazon's own server, free, generally reachable with Ads API credentials, and it creates campaigns and moves bids today. Nova write mode is still in beta.
Closest alternative: PPC Assist and DataDoe also write, both third-party, both paid.
Free and self-hosted
No licence, no vendor, full control. You get raw endpoints and you maintain them.
Closest alternative: Amazon Ads MCP is also free, but scoped to advertising only.
Nova's entry is verified first-hand by introspecting the live server. Everything else comes from vendor documentation, the official MCP Registry or public repository metadata, and each entry says which. Nothing here was tested on a live competitor account.
The data + action MCP for Amazon sellers
Profit you cannot get elsewhere. Net profit per SKU after every Amazon deduction, reconciled to the cent across 40+fee types, with your own COGS and FBM costs inside the number. Every other server on this list either stops at account level or stops at what Amazon publishes.
Search Query Performance in the chat. De-duplicated, with like-for-like period comparisons. Ask which keyword lost share and get the answer, not a report to download.
Write actions with profit context, in beta. Amazon's server can move a bid but does not know your landed cost, so it cannot tell a profitable ACoS from an unprofitable one. A profit tool that cannot act leaves you doing the work by hand. Nova is building the combination, and write mode is in beta today, not shipped.
Your cost data, maintained by talking. COGS per SKU, FBM rates and VAT categories can be set and corrected through the MCP itself, in conversation. That is live now, and no competitor on this list offers it.
One call, not eight. Pre-aggregated, structured tables rather than raw payloads, so a question costs one low-token call instead of paging through orders, settlements and fees and joining them in the model.
The widest modeled Amazon layer on the market: SKU-level profit after every fee, PPC, search query performance, inventory, inbound and 3PL, listing health and BSR, joined in one model, plus live Vendor Central (1P) connections. Write mode is in beta.
Best for: Brands and agencies that want an assistant to answer any Amazon question, from which SKU is losing money after fees to which keyword is worth the bid.
Weakness: Write mode is in beta rather than generally available.
See the 43 tools and what each one returns
Source: Live tool and metric introspection, plus novadata.io documentation. Verified 30 August 2026.
A read layer across nine data domains, including sales, profit, ads, inventory, orders, returns, shipments and keywords.
Best for: Sellers who want a single flat price with operations coverage included on every plan.
Weakness: Fee modeling depth and aggregation guards are not published, so you cannot tell in advance how a margin figure is built.
Source: fluvi.com/mcp. Verified 30 August 2026.
A large tool surface across keyword research, reverse ASIN, product research, rank tracking and listing analysis, plus an account-level profit breakdown.
Best for: Keyword and product research inside an assistant.
Weakness: The profit breakdown is account level, not SKU level, so it cannot answer which product is actually losing money after fees.
Source: Vendor documentation and public repository, July 2026. Verified 30 August 2026.
Read and write across Seller Central, Vendor Central and Ads, with previews and approvals gating every write.
Best for: Teams that want the assistant to execute, and 1P sellers who need Vendor Central coverage.
Weakness: Tables refresh daily, so intraday questions read yesterday's account.
Source: datadoe.com MCP documentation. Verified 30 August 2026.
Read-only market research across products, brands, sellers, subcategories, search terms and ad spy.
Best for: Wholesale and online arbitrage sourcing, brand and category mapping.
Weakness: It reads the market, not your account, so it cannot answer a profit or inventory question.
Source: smartscout.com MCP announcement and partner documentation. Verified 30 August 2026.
Amazon's own server for advertising: campaigns, Sponsored Products, Sponsored Brands, Sponsored Display, DSP and AMC, including creation and bid management.
Best for: Executing advertising changes from an assistant.
Weakness: Scoped to advertising. It cannot see inventory, real margin after fees, Buy Box state or organic rank, so an ads-only agent will happily push spend into a SKU that is unprofitable or about to stock out.
Source: advertising.amazon.com open beta announcement, February 2026. Verified 30 August 2026.
Reported to be in closed beta as of August 2026. No public endpoint or documentation is available, so nothing here is stated as fact.
Best for: Not assessable yet.
Weakness: Nothing published to evaluate. If this is wrong, tell us and we will update the page.
Source: Single unverified public post, August 2026. Verified 30 August 2026.
Servers a seller will find when searching, read from vendor pages, the official MCP Registry or GitHub. None were tested first-hand, and each entry says where its facts came from.
A hosted Seller Central and Amazon Ads server, published in the official MCP Registry under to.agentcentral/amazon-seller-central. The vendor page describes scoped access to Ads, inventory, orders, catalog, rankings, finance and fulfillment after Amazon OAuth, with retained history.
Best for: Sellers and agencies who want a hosted Seller Central connection without running anything.
Weakness: Fee reconciliation depth and refresh interval are not published, so you cannot tell how a margin figure is built. Not tested first-hand.
Source: Official MCP Registry entry and agentcentral.to vendor page. Verified 7 September 2026.
An Amazon Ads and Seller Central server. The vendor page states 50+ tools exposed to Claude, ChatGPT, Cursor and other MCP clients, covering reading performance, editing strategies and rules, proposing changes and executing them.
Best for: Advertising teams who want the assistant to propose and then execute campaign changes.
Weakness: Built around advertising workflows. SKU-level profit after every Amazon fee is not published as part of the model. Not tested first-hand.
Source: ppcassist.com/solutions/mcp vendor page. Verified 7 September 2026.
A read server over Amazon account data, positioned around replacing report and spreadsheet work. The vendor page gives Buy Box share loss by ASIN as an example question.
Best for: Existing Seller Labs customers who want their account data inside an assistant.
Weakness: Tool count, refresh interval, marketplace coverage and fee modelling depth are not published. Not tested first-hand.
Source: sellerlabs.com/amazon-mcp vendor page. Verified 7 September 2026.
Zapier exposes its Amazon Seller Central connector over MCP, so an assistant can trigger the same Seller Central actions a Zap can. It is an automation bridge rather than an analytics model.
Best for: Teams already running Zapier who want an assistant to fire existing Seller Central actions.
Weakness: No modelled P&L, no fee attribution, no analytics layer. You get the actions the connector supports, nothing more. Not tested first-hand.
Source: zapier.com/mcp/amazon-seller-central product page. Verified 7 September 2026.
A reporting platform whose MCP server covers Amazon Seller Central alongside 30+ other marketing sources. A separate Seller Central specific server is published on GitHub, created April 2026.
Best for: Operators who want Amazon sitting next to ad platforms and GA4 in one reporting connection.
Weakness: Amazon is one source among many rather than the modelled centre, and the Seller Central repository has almost no public traction. Not tested first-hand.
Source: twominutereports.com/mcp, mcpservers.org listing and the public GitHub repository. Verified 7 September 2026.
Community repositories that wrap Selling Partner API endpoints as MCP tools. The most visible, mattcoatsworth/AmazonSeller-mcp-server, had 43 stars and was last pushed in April 2025 when we checked.
Best for: Developers who want a free, self-hosted starting point and are happy to maintain it.
Weakness: Unmaintained, no profit modelling, no aggregation guards, and you carry the SP-API credentials, rate limits and upgrades yourself.
Source: GitHub repository metadata, read via the GitHub API. Verified 7 September 2026.
Checked and left out: SellerPPC. It is a real Amazon PPC automation product, but we found no MCP server page or documentation for it, so it does not belong on a list of servers you can connect. If that changes, tell us and we will add it.
Nova MCP exposes SKU-level profit after fees, PPC and search terms, inventory, inbound and listing health across 23 marketplaces, and write mode (in beta) lets the assistant act on them. Connect Seller Central once.
The Model Context Protocol is an open standard for connecting AI assistants to external systems. A server publishes a set of tools with typed inputs; the assistant reads the tool list, decides which one answers the question in front of it, calls it, and reasons on what comes back. The specification and a reference server collection are public on the protocol documentation site and in the modelcontextprotocol/servers repository, which is where most of the ecosystem started.
The important consequence for operators: the assistant is no longer limited by what you remembered to paste into the chat. It is limited by what the server exposes, and by how usable that data is once it arrives.
Every ecommerce MCP server falls into one of two jobs, and confusing them is the most common reason a first connection disappoints.
Until now you needed one of each. Nova closes that gap: it is the widest data server on the market, and write mode, in beta today, turns the same connection into an action server. That is the trajectory that puts an Amazon business on autopilot: one assistant, one connection, answers and execution on the strength of the data model underneath.
Multi-channel brands often add a platform server too, such as the storefront server documented in the Shopify developer docs, which is why two or three connections enabled at once is normal.
Every cell reflects what the vendor publishes, or what we saw by connecting. Last verified 30 August 2026. Where a vendor publishes nothing, the cell says so instead of guessing.
Nova
Source: Live tool and metric introspection, plus novadata.io documentation. Verified 30 August 2026.
Fluvi
Source: fluvi.com/mcp. Verified 30 August 2026.
Helium 10
Source: Vendor documentation and public repository, July 2026. Verified 30 August 2026.
DataDoe
Source: datadoe.com MCP documentation. Verified 30 August 2026.
SmartScout
Source: smartscout.com MCP announcement and partner documentation. Verified 30 August 2026.
Amazon Ads MCP
Source: advertising.amazon.com open beta announcement, February 2026. Verified 30 August 2026.
Sellerboard
Source: Single unverified public post, August 2026. Verified 30 August 2026.
The gap nobody fills by accident
Partly, since 23 September 2026. Amazon publishes an official MCP server for advertising, in open beta since February 2026, and it can create campaigns and move bids. It also launched a Selling Partner plugin, live in Amazon Quick for US-store sellers and in beta with Claude (US stores, waitlist, Professional selling plan), which Amazon says pulls listings, inventory levels and sales analytics. Amazon has not said the plugin covers stores outside the US, and it does not hold your cost of goods, freight, duties or VAT. Profit after those costs still comes from a third party or a wrapper you host yourself.
That is the whole shape of this market. The official routes can see your Amazon account but not your costs. The third-party servers can see, and only some can act.
There are free options, and they are honest about what they are. Community repositories wrap Selling Partner API endpoints as MCP tools. The most visible one, mattcoatsworth/AmazonSeller-mcp-server, had 43 stars and was last pushed in April 2025 when we checked its GitHub metadata on 7 September 2026. Two Minute Reports also publishes a Seller Central repository, created April 2026.
What you get: raw endpoints, no licence fee, full control. What you carry: SP-API credentials, rate limits, breaking changes, and the fact that none of these model profit. An unmaintained wrapper is a fine weekend project and a poor foundation for a decision about restocking.
Amazon's Ads MCP server is the other genuinely free option, and it is maintained by Amazon. It is scoped to advertising.
Multi-channel brands usually add a platform server next to the Amazon one. Two or three connections enabled at once is normal, and they answer different questions: the platform server knows the storefront, the Amazon server knows the marketplace economics. Each of these was checked against its own documentation.
Storefront MCP
ShopifyCatalog search, cart operations and storefront context for agents shopping a Shopify store.
Gap: Storefront-facing. Not a merchant finance or margin layer.
Stripe MCP
StripePayments objects: customers, invoices, subscriptions, balance and payment intents, plus documentation search.
Gap: Payments only. No view of marketplace fees or fulfilment cost.
PayPal MCP
PayPalInvoices, orders, subscriptions, disputes and transaction reporting on the PayPal account.
Gap: One payment method's view of revenue, not the whole store.
Klaviyo MCP
KlaviyoProfiles, lists and segments, campaigns, flows and campaign performance metrics.
Gap: Marketing engagement, not order economics.
AWS MCP servers
AWS LabsA large open-source collection covering AWS services: documentation, cost analysis, databases, infrastructure.
Gap: Infrastructure, not commerce. Nothing here reads a Seller Central account.
Some Amazon figures must never be summed. Search query volume repeats on every parent ASIN that ranks on it. FBA quantities share pools across SKUs. A good server refuses those additions in the tool contract rather than trusting the model to know. Four guards built into Nova:
Search volume, de-duplicated
Naively summing search query rows overstates demand by roughly 2.4x. Nova returns the de-duplicated total.
Like-for-like comparisons
Year on year is computed only over keywords present in both periods. Raw totals can read +34% where the truth is -3%.
Calendars aligned
Amazon search weeks start Sunday, ad weeks start Monday. Nova states the exact days each period covers.
Pools and namespaces separated
Shared FBA pools are flagged, and product, ads and search metrics stay in separate universes so nothing cross-sums.
Five questions separate a server you still use in three months from one you disable after a week.
Does it join the data, or hand over silos?
Profit is orders minus every Amazon fee minus your landed cost minus ad spend. A server that returns each of those from a different endpoint pushes the join into the model, which is exactly where it goes wrong.
Does it stop the assistant computing a wrong number?
Some Amazon figures must never be summed. Search query volume repeats on every parent ASIN that ranks on it. FBA quantities share pools across SKUs. A good server refuses those additions in the tool contract rather than trusting the model to know.
How fresh is the answer?
Daily tables are fine for a monthly review and useless for a stock-out call at 4pm. Ask what the refresh interval is on money metrics specifically, not on the dataset as a whole.
Read, or read and write?
A read server informs the decision. An action server executes it. They are not competing products, and most people running assistants daily end up with one of each.
Auth and auditability
OAuth with revocable per-account credentials is the baseline. Beyond that, ask whether calls are logged with the reason they were made. Anyone handling client data will be asked this eventually.
Cost is the one people underestimate. A thin API wrapper often needs eight to twenty calls, because the assistant has to page through orders, then settlements, then fees, then join them itself. Amazon also enforces per-operation rate limits, documented in the SP-API usage plans reference, so those calls get slower exactly when the date range gets interesting.
What each category of ecommerce MCP server actually covers
| Server type | Typical scope | Read or write | Best for |
|---|---|---|---|
| Nova (data + action) | P&L, fees, PPC, search terms, SQP, inventory, inbound and 3PL, listings, BSR | Read, write in beta | Profit, pricing and inventory decisions, then execution |
| Amazon Ads MCP (first-party) | Sponsored ads campaigns, bids, reporting | Both | Executing ad changes from chat |
| Shopify MCP (first-party) | Storefront catalog, cart, store data | Read plus commerce actions | Shopify-native storefront workflows |
| Research servers | Brands, sellers, categories, competitor keywords | Read | Sourcing and market mapping |
| Community SP-API wrappers | Raw Selling Partner API endpoints | Read, sometimes write | Developer experiments |
Find yourself in the left column. Where a competitor is the right answer, it says so.
Private label brand, 7 to 8 figures
Nova, the data + action MCP
The question you ask most is which product is actually making money after fees, and then what to do about it. Nova answers it from one joined data model, and write mode, in beta, lets the assistant carry out the decision too.
Agency or aggregator, many client accounts
Nova
One connection covers every client account, scoped per client, with each call recorded alongside the reason it was made. That is the part a client will ask about.
Wholesale or online arbitrage
SmartScout, plus Nova once your own catalogue matters
Your first question is about the market, not your P&L. Use a research server for sourcing, and add a data server when margin per SKU starts driving decisions.
Vendor Central (1P)
DataDoe
DataDoe covers Vendor Central today. Nova Vendor Central support is live, so you can run both 1P and 3P in one workspace.
Ads-only operator
Amazon Ads MCP, plus Nova as the guard rail
An ads-only agent cannot see margin, stock cover or Buy Box state, so it will bid confidently on a SKU that loses money or is about to run out. The read layer is what stops that.
Our take
Connect the server that carries your own numbers and can act on them. Nova reads the whole business, and write mode, in beta, lets the assistant execute as well.
Best fit if
Skip if
Two things are changing fast. First, client support is converging: remote MCP servers now work across the major assistants, so vendors ship one endpoint rather than one integration per model. Second, the competition is moving from "do you have an MCP" to "how good is your data model", because every vendor can expose tools and very few can expose a reconciled P&L. Read and write are converging too: the servers that win are the ones that answer the question and then carry out the decision.
For sellers, that is the useful filter. The assistant is a commodity. The data underneath it is not. Related reading: Amazon Ads MCP compared with Nova MCP, Amazon MCP for agencies and the Claude setup walkthrough.
Test it in 5 minutes
You get rows: SKU, profit change, and the fee, price or ad-spend line that moved it.
Start the free trialWhat operators ask before connecting the first server
Page last updated 7 September 2026. Nova's entry is verified first-hand by introspecting the live server. Other entries come from vendor documentation, the official MCP Registry and public repository metadata, never from testing a competitor account. Where a vendor publishes no figure, the cell says "not published" rather than an estimate. Reviewed quarterly, and corrections from vendors are welcome.
Join thousands of successful sellers who use Nova Analytics to make data-driven decisions and maximize their profits.
Explore more expert insights to grow your Amazon business
Claude MCP finally lets AI answer real questions about your Amazon business. This guide covers what MCP is, what the Amazon MCP landscape looks like in mid-2026, and how to connect Claude to your live Nova data in an afternoon.
Wrapping an API in MCP takes an afternoon. Shipping an MCP a business can run on is everything between your database and the model's first token: context economics, semantics, reconciliation, guardrails and observability.
Once your Amazon data is reachable through an MCP connection, the bottleneck moves from access to asking. These are the prompts that change what a seller does on a Monday, with the phrasing that keeps answers accurate.
