Claude MCP for Amazon Sellers - Live Data in One Query 2026

Claude MCP finally makes it easy to ask Claude real questions about your Amazon business. But the answer is only as good as the data model Claude reads. This guide covers what MCP is, what Nova MCP exposes, and how to point Claude at your live seller data in an afternoon.

What Claude MCP actually is

Model Context Protocol (MCP) is the open standard Anthropic published in late 2024 to let AI assistants call external tools through a structured interface. Instead of pasting CSVs into a chat, you run (or connect to) an MCP server that exposes tools. Claude decides which tools to call, calls them, and reasons on the returned data inside the same conversation.

If you want the reference material, Anthropic keeps the announcement and the open spec live at anthropic.com/news/model-context-protocol and the tooling documentation lives at modelcontextprotocol.io.

For Amazon sellers, the important thing is that MCP is a data protocol, not a data model. Anyone can publish a server; the quality of the answers Claude gives depends on what that server actually joins together for it.

The Amazon MCP landscape in mid-2026

Three groups of servers matter today:

• Amazon Ads MCP Server - open beta, covers Sponsored Products, Brands and Display metrics from Amazon's ad accounts. Announcement at advertising.amazon.com. Useful, but ads-only.
• Community SP-API wrappers - GitHub projects that thinly wrap Selling Partner API endpoints. They give Claude a socket to SP-API but no data model: it still has to reconcile fees, joins and rate limits inside the prompt.
• Nova MCP - a seller-grade MCP server that exposes Nova's joined Amazon data model (orders, 40+ fee types, FBA inventory, PPC at product level, organic) as tools Claude can call in one turn.

Pick a server based on the question you want Claude to answer. If it is "what happened in my ads this week", Amazon's own MCP is fine. If it is "which 10 SKUs killed my margin across DE and UK last week and why", you need a seller-grade model, not a raw API socket.

What Nova MCP exposes to Claude

Nova MCP is a read-only Claude connector on top of the same Nova data model that powers the dashboard. Everything in the model is joined at the SKU and marketplace grain, refreshed hourly, and covers all 21 Amazon marketplaces.

Concretely, Claude can query:

• Reconciled SKU profit across 40+ Amazon fee types
• FBA inventory, inbound and days of cover per ASIN
• Sponsored Products, Brands and Display spend at the product level
• Organic share, Buy Box, Subscribe & Save, BSR on your own ASINs
• Settlement-grade P&L per marketplace and per date range

The tools are read-only. Claude can query but never write. Credentials are per account and revocable at any time from your Nova settings.

Connecting Claude to Nova MCP

The full walkthrough with screenshots lives in the connect-Claude tutorial:

The short version:

1. Sign in to Nova and connect Seller Central and Ads accounts if you have not already.
2. In Nova, open the MCP panel and generate a read-only credential for Claude.
3. Add the endpoint to Claude's MCP settings. Save.
4. Ask Claude a real question. The first working answer usually lands inside 24 hours.

Prompts that work well on day one

These are the prompts our team runs when we onboard a new account. They exercise the model across profit, inventory and ads without needing you to know any SQL.

• "Rank the 10 SKUs that lost the most contribution margin in the last 30 days across all marketplaces. For each, cite the top driver: fees, refunds, or ad spend."
• "Draft a weekly P&L brief with the 5 biggest revenue and profit swings by marketplace, with root cause per row."
• "Which ASINs will stock out in the next 21 days? Return days of cover and recommended reorder quantity."
• "Which Sponsored Products campaigns spent above the break-even ACoS last week? Show ACoS, spend, and sales share."
• "Compare DE and UK contribution margin this quarter vs last. Highlight the SKUs responsible for the delta."

The point is not the prompt list; it is that Claude answers each of these in one turn because Nova already joined the underlying tables. Feed the same prompt to a raw SP-API server and you will burn 8 to 20 tool calls chasing pagination and reconciliation.

From Q&A to agents

Once Claude is answering ad-hoc questions well, the natural next step is to turn the best prompts into repeatable agents. Nova customers ship three patterns most often:

• Profit Watchdog - a daily Claude run that flags any SKU whose contribution margin dropped more than a threshold and posts the summary to Slack.
• Restock Copilot - a weekly run that returns the reorder list with days of cover and lead-time buffer.
• Ad Spend Auditor - a run after each ad-spend batch that names the campaigns above break-even ACoS with the exact reason.

None of these need custom code. They are prompts scheduled against a Claude client that has Nova MCP wired in.

Where Nova MCP stops

Two things Nova MCP is deliberately not:

• It is not an action layer. Claude reads and reasons; if you want it to push bids or edit listings, pair Nova MCP with Amazon's own Ads MCP or your action tooling.
• It is not a keyword-research surface. For SEO or keyword scouting, use a research tool that crawls the SERP. Nova is about your own numbers, not the wider marketplace.

Everything else that reads from your Amazon business - profit, ads, inventory, organic on your own ASINs, cross-marketplace briefings - is what Nova MCP is built for.