200+Data Points

Hourly Refresh

21Marketplaces

100%Normalized Data

99.9%Uptime SLA

Amazon MCP for Sellers

What is an Amazon MCP server?

MCP (Model Context Protocol) is the open standard that lets an AI assistant call external tools. An Amazon MCP server exposes your seller data as tools the assistant can invoke mid-conversation. The difference between one server and another is what it can see, and whether the numbers it returns are reconciled.

Try for Free

Rated 4.7/5 onTrustpilotfrom 23 reviews

What is Amazon MCP?

Amazon MCP is the use of the Model Context Protocol to let an AI assistant query Amazon seller data directly. An MCP server exposes that data as tools the assistant calls inside a conversation. Amazon's own MCP server covers advertising. Nova MCP covers the whole business: SKU-level profit across 40+ fee types, Sponsored Products down to the search term, FBA and AWD inventory, listing health and BSR, in 21 marketplaces, refreshed hourly and read-only.

How the read layer works

From fragmented Amazon APIs to any AI assistant, in one flow

Nova absorbs SP-API, Ads, and Reports complexity. Your AI reads one clean schema.

Step 1

Amazon APIs

Fragmented, schema-shifting

SP-APIrate-limited

Ads APIshifting schemas

Reports APIlags 24 to 48h

We absorb the chaos so Your AI does not have to.

Step 2

One schema, pre-modeled for AI

Sales & Profit

Net Profit at SKU levelOrdersFeesCOGSSettlementsReturns

Advertising & Discoverability

PPC KeywordsSearch TermsSessionsBuy BoxBSR

Catalog & Inventory

FBA InventoryListingsCatalogueSubscribe & Save

200+

metrics

marketplaces

SKU

level grain

Hourly

refresh

Step 3

Your AI

Reads Nova in one query. Returns profit, fees, ads, inventory answers in seconds.

Live, hourly refresh

What an Amazon MCP server actually is

MCP (Model Context Protocol) is the open standard for letting AI assistants call tools. An Amazon MCP server wraps Amazon data and exposes it as callable functions. Nova MCP is live today for Claude, ChatGPT, Gemini and any other MCP-ready client.

Amazon ships an official Ads MCP server. The community publishes SP-API wrappers. They all do the same thing: hand the AI a phone line to Amazon. What they do not hand it is a seller-grade data model.

That gap is the difference between "the assistant tried to compute my margin and timed out" and "the assistant told me the 5 SKUs that lost money this week and why."

The problem with raw MCP

Why "give the AI SP-API" rarely works

Token bills explode

8 to 20 tool calls per question, with pagination and retries. One question can burn dollars in tokens.

No profit model

SP-API returns orders, fees and settlements in different shapes. The AI tries to reconcile them in the prompt. It gets the answer wrong.

Schema fragility

Amazon ships breaking changes regularly. Every change breaks your AI workflows until someone patches the wrapper.

Amazon Ads MCP vs raw SP-API wrappers vs Nova MCP

Same assistant, same prompt, three very different context layers.

Dimension

Amazon Ads MCP

Raw SP-API wrapper

Nova MCP

Scope

Ad account only

Raw SP-API endpoints

Full seller model: P&L, ads, inventory, listings, BSR

SKU-level profit

Not available

AI must reconcile in the prompt

Reconciled across 40+ fee types

Tool calls per question

Several, async reporting

8 to 20 with pagination

One call, filtered server side

Token cost per answer

Medium

High, scales with retries

Low, pre-computed metrics

Freshness

Async report jobs

Per call, rate limited

Hourly refresh

Marketplaces

Per ad account

One auth per region

21 marketplaces, one model

Writes

Yes, campaign actions

Depends on the wrapper

No, read-only by design

Best fit

Executing ad changes

Developer experiments

Deciding: profit, planning, alerts

The tools Nova MCP exposes to your assistant

Nine read-only tools. Each one answers a class of seller question in a single call.

query_metrics

Any metric over any period, grouped by SKU, ASIN, marketplace, account or a time grain. Ranking and thresholds run server side.

find_movers

The SKUs whose value moved most between a period and a baseline, in either direction. This is the 'what changed' tool.

explain_movement

For one SKU, the causal bundle: traffic, conversion, price, buy box, promotions, ads and rank versus a baseline.

query_advertising

Sponsored Products at campaign, ad group, targeting type, keyword, attributed ASIN or SKU, plus any time grain. 13 KPIs per row.

query_search_terms

The customer queries behind the spend. Negative-keyword candidates and auto-campaign discoveries, on demand.

query_inventory

FBA and AWD stock health: available, inbound, reserved, sales velocity, days of inventory, out-of-stock risk.

query_listing_health

Current listing state: suppressed, removed, not buyable, buy box lost, hijackers, plus published catalog content and backend keywords.

query_bsr

Best Seller Rank for your own ASINs, latest or as a time series, with the category and rank pair kept intact.

list_cost_inputs

The per-SKU COGS, merchant-fulfilment cost and VAT category you configured, so the assistant can audit your cost setup.

What sellers actually ask

Three real questions, and the tools behind each answer.

"Which keywords burned more than $100 with zero sales last month?"

One query_search_terms call with a spend threshold and a zero-sales filter. The answer comes back as a ranked table: query, campaign, match type, spend, clicks. That is a negative-keyword list, not a report.

"Which SKUs lost the most net profit versus last month, and why?"

find_movers ranks the losers, then explain_movement pulls the causal signals for the worst one: price change, buy box share, ad spend, conversion rate, rank.

"What runs out of stock in the next three weeks across all marketplaces?"

query_inventory sorted by days of inventory, with AWD stock counted in coverage. The assistant returns the reorder shortlist with velocity and projected cover.

Use both. MCP for actions. Nova for answers.

The smart stack pairs an MCP action layer with Nova as the read layer.

MCP action layer

Use Amazon's Ads MCP and SP-API wrappers to push changes: bid updates, listing edits, inbound creation.

Good at: writes, mutations, single-record actions.

Nova read layer

Point your AI at Nova for everything it needs to reason: profit, fees, ads efficiency, inventory risk, listing health, organic share.

Good at: analytics, planning, alerting, briefings, agents.

Live in 24 hours, not 24 weeks

Four steps from Amazon to your assistant via Nova MCP.

Step 01

Connect Amazon

Auth your Seller Central and Ads accounts in Nova. 21 marketplaces supported.

Step 02

Let Nova model it

Orders, settlements, 40+ fee types, ads and inventory are joined at SKU level and refreshed hourly.

Step 03

Add the MCP endpoint

Point Claude, ChatGPT or Gemini at Nova MCP. Read-only, per-account credentials, revocable anytime.

Step 04

Ask anything

Profit, wasted ad spend, stockout risk, suppressed listings. One question, one call.

See the step-by-step Claude setup

Everything AI + Nova

How to connect Claude Amazon Data for Claude Amazon AI Agents Every metric Nova exposes Claude MCP for Amazon guide

About Amazon MCP for Sellers

Common questions about MCP, SP-API, and Nova as the seller read layer

MCP (Model Context Protocol) is the open standard that lets an AI assistant call external tools. An Amazon MCP server exposes Amazon seller data as tools the assistant can invoke inside a conversation. Amazon ships an official Ads MCP server for the advertising account. Nova MCP exposes the full seller data model: profit, 40+ fee types, PPC, search terms, inventory, listing health and BSR across 21 marketplaces.

Not for the full Selling Partner API. Amazon's official MCP server covers Amazon Ads. AWS publishes MCP servers for AWS services, and the community publishes thin SP-API wrappers. None of them return a reconciled seller P&L, so none of them can answer a margin question on their own.

A raw wrapper gives the AI a socket, not an answer. To compute SKU profit the assistant has to call orders, then settlements, then fees, then ads, then reconcile them by SKU inside the prompt. That is 8 to 20 tool calls, a large token bill, and usually a wrong number, because the fee logic and the joins are not in the model.

Profit questions (which SKUs lost margin last week and why), PPC questions (which keywords burned over $100 with zero sales last month), inventory questions (which ASINs run out inside 21 days), listing questions (which listings are suppressed or losing the buy box right now), and cross-marketplace briefings. Each of those is one tool call, filtered and ranked server side.

query_metrics for any metric at any grain, find_movers for biggest gainers and losers, explain_movement for the causal signal bundle behind one SKU, query_advertising for Sponsored Products at campaign, ad group and keyword grain, query_search_terms for the actual customer queries, query_inventory for FBA and AWD stock health, query_listing_health for current listing state and buy box, query_bsr for Best Seller Rank, and list_cost_inputs for the per-SKU COGS and VAT setup.

Any MCP-capable client. Claude (Desktop, Code, web), ChatGPT, Gemini and other assistants that support remote MCP servers. You bring your own AI account; Nova provides the endpoint and the read-only credentials.

Yes. The assistant reads the joined Nova model through per-account credentials you can revoke at any time. It cannot write to Seller Central or to your ad account. Action layers such as bid pushes or listing edits stay with your team or with Amazon's own Ads MCP.

Yes, and many teams do. Amazon Ads MCP executes changes on the ad account. Nova MCP reads and reasons across the whole business: profit, fees, inventory, listings, organic. One executes, the other informs, and both can sit in the same assistant.

Metrics are pre-computed and returned as structured tables, filtering and ranking happen server side, and one question maps to one call instead of a pagination loop. The assistant reads a small ranked result set rather than thousands of raw rows.

Nova refreshes hourly, so the assistant answers on today's numbers rather than last week's export. Advertising and search-term data follow Amazon's own reporting cadence.

All 21 Amazon marketplaces Nova supports. You can ask a consolidated question or a per-marketplace one from the same model.

Credentials are per account and read-only, scoped to your own seller data, and revocable from Nova at any time. Nothing is written back to Amazon, and the assistant never receives your Seller Central login.

No. Connect Seller Central and Ads in Nova, add the Nova MCP endpoint in your assistant, approve the connection. Most teams ship their first working prompt the same day.

No. Dashboards answer the questions you planned for, the same way every morning, for the whole team. The assistant answers the one-off question you never built a view for. Same data model underneath, two surfaces.

The rest of the Amazon MCP guide

Every page in this cluster answers a different question. Start at the pillar if you want the full picture.

Give your AI seller-grade Amazon data

Live in 24 hours. Profit-accurate. 21 marketplaces.