Before any code: six accounts, a $0 stack, and three gotchas
I set up the full ShelfReady stack on free tiers before writing a line of app code, and hit three surprises that would have cost me the first episode.
The goal. ShelfReady answers one question for Shopify merchants: can an AI shopping agent actually find and buy your products? Before building anything, I wanted every service connected and tested, so Episode 01 could start on the product, not on account settings.
The product call. I set a hard constraint: $0 a month until the product proves its value. That meant Vercel Hobby, Neon's free Postgres, Upstash's free Redis, and Gemini's free tier as the default model. The trade-off is rate limits and a model that may not be the best one. So I logged a decision: build on Gemini now, and compare models later with the same eval set, once there's a baseline to compare against.
What I set up.
- A GitHub repo and a Vercel project, ready to deploy on first push
- A Shopify development store and a custom app with the scopes the product needs
- Neon Postgres for the catalog copy and eval results; Upstash Redis for rate limiting
- A Gemini API key, tested with a real generation call
- A written spec, a scoring rubric and a decisions log, before any code
What broke (or almost did).
- Shopify's bare developer URL opens an internal employee login. The right place is the Dev Dashboard.
create-next-apprefuses to run in a folder that already has files, and my starter kit was already there. Fix: move the kit aside, scaffold, move it back.- Google AI Studio now issues keys in a new format. Generation calls worked, but the list-models endpoint returned 403. So the status page only checks that the key exists instead of calling an endpoint that fails for reasons unrelated to the app.
The number. 6 services connected and verified the night before. Running cost: $0 a month.
Next. Episode 01: scaffold the app, connect Shopify and the database, and get a live status page that's all green.