Amazon Seller MCP

A self-hosted analytics brain for one Amazon seller account. It pulls your Seller Central + Advertising data on a schedule into your own Postgres, then serves it to Claude (or any MCP client) over a bearer-token-guarded HTTP endpoint — so you can ask "where am I wasting ad spend?" and get an answer grounded in your real numbers.

Single-tenant by design. One deployment = one Amazon account = your data, your Railway project, your secrets. Nobody else authorizes your apps and no data leaves your infrastructure.

Quickstart

git clone https://github.com/stevegustafson32/amazon-seller-mcp && cd amazon-seller-mcp
cp .env.example .env                              # fill in your own credentials
docker compose up -d db                            # local Postgres (Railway provides its own)
export DATABASE_URL=postgresql://postgres:postgres@localhost:5432/amazon_seller_mcp
python -m pytest                                   # seeds fixtures, verifies all 7 tools

Deploying for real (non-technical)? Follow DEPLOY_GUIDE.md — it walks you click-by-click from zero to a live MCP your Claude can connect to.

One click provisions all three components — Postgres + worker + mcp — and prompts you for every credential at deploy time (nothing is baked into the template). Prefer a guided walkthrough? Follow the click-by-click DEPLOY_GUIDE.md.

Related MCP server: Shopify Partner Agent

How it works

Your Amazon account            Your Railway project                  Your Claude / Cowork
-------------------            --------------------                  --------------------
SP-API app  (self-authorized) ─┐   ┌─ worker (cron ingestion) ─┐
Ads API app (refresh token)   ─┼─► │  Postgres (your data)     │ ─► MCP web service ─► custom
Data Kiosk / Pricing          ─┘   └─ writes normalized tables ┘     (HTTP + bearer)     connector

Claude never calls Amazon live. The worker owns every Amazon call — auth, pagination, throttling, retries — and writes normalized tables. The MCP web service only reads Postgres, except the single approval-gated write path (propose → confirm) that changes a bid.

Tools your Claude gets

Read

• query_ads_performance — ACOS / ROAS / CTR / CVR by campaign, target, or search term

• find_search_term_waste — spend with zero conversions, ranked

• find_keyword_harvest — converting search terms to promote to exact keywords

• find_cannibalization — terms you win organically yet still pay to advertise (Brand Registry)

• get_sqp_share — Search Query Performance share (Brand Registry)

• get_buybox_status — who owns the Buy Box, and whether it's you

• get_pnl_snapshot — sales, ad spend, ACOS, TACOS vs. break-even per ASIN

Write (approval-gated)

• propose_bid_change → returns a diff and a one-time token. Writes nothing.

• confirm_bid_change → applies the change via Ads API, only with that exact token.

Bookmarkable dashboard — GET /dashboard?token=<MCP_BEARER_TOKEN> server-renders a single self-contained page (KPI row, benchmark scorecard, 8-week trends, top books, negative/harvest lists, Buy Box, inventory health) from your Postgres. Thresholds, benchmarks, and which panels show all come from config/profile.yaml. Add ?range=<days> to change the KPI window. Bookmark {APP_URL}/dashboard?token=….

Layout

amazon_seller_mcp/
  shared/      config.py (env validation), db.py (pool + migrations), lwa.py (token exchange)
  ingestion/   sp_reports, sp_datakiosk, sp_pricing, ads_reports, ads_entities, scheduler
  mcp/         server.py (remote MCP + bearer auth), tools.py, metrics.py, ads_auth.py (hosted OAuth)
db/            migrations/ (idempotent DDL), views.sql (metric layer), fixtures/ (seed data)
config/        profile.yaml (business defaults + Good/Watch/Fix thresholds)
scripts/       check_setup.py, get_ads_token.py, backfill.py
tests/         all 7 read tools + write path + auth + profile, against fixtures

Note on the package name. The build spec named the server folder mcp/, which would shadow the official mcp SDK package on import. Everything is nested under the amazon_seller_mcp package so every spec folder name is preserved while the SDK stays importable. Commands are therefore python -m amazon_seller_mcp.mcp.server and python -m amazon_seller_mcp.ingestion.scheduler.

Configure your business — config/profile.yaml

The analytics layer reads defaults and thresholds from config/profile.yaml instead of hard-coding them. It ships with sensible Books/seller-publisher defaults so everything runs before you touch it. Edit it to set your business_model, unit economics (which compute your default break-even ACOS and target ACOS), the benchmark set, per-metric Good / Watch / Fix thresholds, and which panels are enabled (turn off the Brand-Registry panels if you're not enrolled). Per-SKU rows in the sku_margins table override the profile's break-even for that SKU.

Security

• The MCP service rejects any request without the correct MCP_BEARER_TOKEN. Tokens are never logged.

• Every account-mutating action goes through propose → human approval → confirm. No silent writes.

• Official Amazon APIs only. No browser automation against Seller Central.

• check_setup.py and get_ads_token.py print secrets to your own terminal only.

Getting the Ads refresh token — two ways

The Ads API needs a browser OAuth round-trip (SP-API doesn't). Either works:

• Hosted (no terminal): once deployed, visit https://YOUR-URL/setup/ads-auth?token=YOUR_MCP_BEARER_TOKEN. It walks you through Amazon consent and renders your ADS_REFRESH_TOKEN + ADS_PROFILE_ID to paste into Railway. The route is bearer-guarded to start; the Amazon callback is authorized by a single-use state it issued.

• Local fallback: python scripts/get_ads_token.py does the same via a localhost callback before you deploy.

Development

python -m pytest needs a Postgres (docker compose up -d db) and TEST_DATABASE_URL/DATABASE_URL pointing at it. Tests seed db/fixtures/ and assert exact results — so the analytics are verifiable before any real Amazon credentials exist. CI (.github/workflows/ci.yml) runs the same on every push.

MIT licensed. See LICENSE.