This isn't a demo. Claude Code has been running this store's ops for a while now — audits, listing edits, review requests, tracking — through the same APIs Amazon gives every seller. $130K sold, a quarter-million dollars of stock under management. 455 review requests on autopilot, 4 broken listings found and patched. Today I'm open-sourcing the whole setup, and publishing the next 30 days of numbers here, win or lose.
git clone https://github.com/EbrahimElbagory/claude-fba.git
Every change ships against a baseline. Current cycle's was captured 2026-07-22; the Day-30 column fills in public as weekly numbers land. No projections, no cherry-picking.
FOLLOW THE NUMBERS ON X →Drop your email — first in line when this becomes a product you can point at your own store. Amazon today; Shopify and eBay decided by demand. No newsletter, no spam.
Which store? Doesn't matter — the numbers do. Every figure below is pulled straight from the seller APIs. Identity withheld so the experiment stays clean; specs published so you can judge whether this scale means anything to yours.
One real session from the log — 2026-07-22, my own store. This is what a routine audit-and-fix cycle looks like when the operator is an agent with API access. Every number came straight out of an API response, nothing is a projection. Caught that session: an estimated $3.5–9K/yr in process leaks.
The market layer runs on Helium 10's official MCP (paid plan). The store layer runs on a free self-authorized SP-API app you create once, in ~30 minutes. SETUP GUIDE →
Keyword volume, ranks, listing quality scores, CPR, competitor intel.
Listings, orders, review solicitations, reports. A ~300-line stdlib client — no pip installs.
Daily review requests 10:00, weekly impact snapshots Mon 09:00.
A full checkup of every listing you own: content quality, which keywords you rank for and where, sales-rank trend, reviews, and months of stock left. Each audit is compared to the last one, so you see exactly what moved.
Amazon lets you request one review per order, day 5–30. This sends that official request for every order you've missed (my catch-up: 437), then runs every morning so no order is missed again. Can't double-send.
Edits live listing text (titles, backend search terms) safely: checks what Amazon allows on your product type, drafts the copy, changes nothing until you approve the exact wording. Old version saved first — undo is one command.
Proves whether changes worked. Records every number before you touch anything, then logs traffic, conversion, units, and revenue per product weekly. You compare against a baseline instead of guessing.
Finds keywords worth going after: what competitors rank for that you don't, expanded seed terms, all sorted by how cheap each is to win. Ends with a map — these words go in your title, bullets, backend, and ads.
The ads loop: pulls every campaign, finds search terms burning money with zero sales, flags winners stuck in broad match, and drafts bid, budget, and negative-keyword changes — nothing spends a cent until you approve it. In testing on my store now.
Letting an agent touch a revenue-producing store only works with hard rules. This store is live cargo — handled accordingly.
A full audit writes nothing. Most sessions never mutate anything.
No buyer-visible change ships until you approve the exact copy in chat.
Pre-edit state saved per SKU; post-PATCH re-read checks for new errors. Rollback is one command.
Your own SP-API app, Helium 10's own MCP. Nothing scraped, no ToS games.
No API path for something? The skills say so and hand you a manual checklist — instead of pretending.
No. It's your own store accessed through your own self-authorized SP-API app — the same official interface every seller tool uses. Review requests go through Amazon's Solicitations API, functionally identical to the "Request a Review" button in Seller Central. Nothing is scraped.
Only for the keyword side — search volumes, rank tracking, competitor keyword data (their official MCP, paid plan). Everything else needs nothing but the free SP-API app you make yourself: review automation, listing edits, impact tracking, sales analytics, the whole loop.
Claude Code (your existing subscription) + optionally Helium 10. The toolkit itself is free, MIT-licensed, stdlib-only Python — no servers, no pip installs. A full audit uses ~25 of your 1,000 monthly Helium 10 MCP calls.
The harness is built so it can't do so silently: schema checked before edits are promised, every change drafted for your approval, pre-edit state backed up before any write, result verified after. Rollback = re-apply the backup.
Scripts and skills run anywhere Python 3.9+ does. The two automation templates are macOS launchd; use cron on Linux or Task Scheduler on Windows.
Amazon Ads API campaign automation (CPR-sized ranking pushes — the auth client already ships in scripts/), Search Query Performance funnel analysis, multi-marketplace support. Issues and PRs welcome.