Amazon FBA Ops MCP Server

A local MCP server that lets Claude query your Amazon Selling Partner API (SP-API) account directly, for two things:

• Profitability — revenue, Amazon fees, reimbursements, cost-of-goods (COGS), and net margin, grouped by SKU / order / day.

• Inventory alerts — FBA stock levels and low-stock warnings based on estimated days-of-supply.

Everything this server does against SP-API is read-only (GET requests). The only writes are to a local JSON file you control (data/sku_costs.json), which stores your own product costs since Amazon has no concept of COGS.

Not built yet (future phases): Amazon Ads/PPC, keyword research, Slack notifications, scheduled/cron runs, a dashboard UI.

1. Get SP-API credentials

You need three values from Amazon: an LWA app ID, an LWA client secret, and a refresh token. No AWS IAM / AWS keys are needed — SP-API dropped that requirement in October 2023.

1. Seller Central → gear icon → Apps and Services → Develop Apps → Add new app client. Choose Private app (not "Publish for others") — this is for your own seller account only.

2. When prompted for API access/roles, select the roles that cover Orders, Finances, Inventory and Order Tracking, and Product Listing (label wording can shift in the Seller Central UI — pick whatever maps to those four areas). Amazon may take a short time to approve the role grant even for a private app.

3. Open the app's detail page and copy the LWA Client ID and LWA Client Secret.

4. Still under Develop Apps, click Authorize on your app (self-authorization — since you're authorizing against the same account that created the app, no OAuth redirect server is needed). Confirm, and copy the refresh token shown — it's shown only once; if you lose it, re-run Authorize to get a new one.

5. Find your marketplace (e.g. US, UK, DE, JP) — this determines both the API endpoint region and the encoded marketplace_id.

Related MCP server: Amazon Seller MCP Server - DataDoe

2. Configure

cp .env.example .env

Fill in .env:

LWA_APP_ID=...
LWA_CLIENT_SECRET=...
SP_API_REFRESH_TOKEN=...
SP_API_DEFAULT_MARKETPLACE=US   # match your seller's marketplace

.env is gitignored — never commit it.

3. Install

python3 -m venv .venv
.venv/bin/pip install -e .

4. Sanity-check credentials before wiring into Claude

PYTHONPATH=src .venv/bin/python3 -c "
from amazon_mcp.orders_service import get_orders
from datetime import datetime, timedelta, timezone
end = datetime.now(timezone.utc)
start = end - timedelta(days=7)
print(get_orders(start.strftime('%Y-%m-%d'), end.strftime('%Y-%m-%d')))
"

If this prints an orders list (even an empty one) instead of an error, your credentials work. If it errors, re-check .env against step 1 before touching the MCP server itself.

5. Set your product costs

Profitability can't be computed without knowing what each SKU costs you to make/source. Use the set_sku_cost tool from Claude once the server is registered, or seed data/sku_costs.json directly (see data/sku_costs.example.json for the shape). SKUs with no configured cost show up in missing_cogs_skus in the profitability output rather than silently being treated as free.

6. Register with Claude Code

A .mcp.json is already checked into this project root. It points at .venv/bin/python, so make sure step 3's install happened inside this directory. Claude Code will pick it up automatically when you open this project — restart Claude Code (or run /mcp to reload) if it doesn't appear right away.

To also register it in Claude Desktop, add the same entry under mcpServers in ~/Library/Application Support/Claude/claude_desktop_config.json.

7. Verify end-to-end

All of these are safe, read-only checks against your real seller account:

• get_orders over a known recent window — cross-check the count/total against Seller Central's own Orders report.

• get_inventory_summary() — cross-check fulfillable quantities against the FBA inventory dashboard for 2-3 known SKUs.

• get_catalog_item(asin) for a known ASIN — confirm title/image match the live listing.

• calculate_profitability over a short window — manually reconstruct expected profit for 1-2 orders (revenue from Seller Central, fees from the Payments view, COGS from what you entered) and compare.

• get_low_stock_alerts — temporarily lower days_of_supply_threshold to confirm it flags a SKU you know is low.

Run the unit tests (mocked, no network) any time after changing code:

.venv/bin/pip install -e ".[dev]"
PYTHONPATH=src .venv/bin/pytest

Tools

Known limitations

• Orders and financial events have a ~48-hour lag on Amazon's side — very recent orders won't show up yet.

• get_low_stock_alerts and calculate_profitability (when grouped by SKU) derive sales velocity/units by calling get_order_items once per order in the window, capped at 200 orders (velocity_sampled / truncated flags tell you if a cap was hit). High-volume sellers will eventually want a Reports-API-based bulk pull instead — not built in this MVP.

• Financial-event field parsing (finances_service.py) is based on Amazon's documented Finances v0 schema but has not yet been checked against a live response — if the shapes don't match exactly, cross-check get_financial_events's raw output the first time you run it for real and adjust _flatten_* in finances_service.py if needed.