Twenty competitor pages in,
you have forgotten what the first one said.
Deciding whether an item is worth doing means holding twenty listings, a supplier quote sheet, a returns policy and a few hundred reviews in your head simultaneously and then having a feeling about it. The feeling is the part you are actually good at. The holding is the part that does not scale, and it is why the third product of the evening gets a worse look than the first one did.
The reading you can hand over
A scoring pass that comes back the same shape every time
Category viability, price bands, the competitor picture, margin assumptions and sourcing risk in a fixed layout, because that is what makes two candidates comparable. A prompt you rewrite each session gives you two reports you cannot lay side by side, which is precisely the thing you needed when you were choosing between them.
The quote sheet against the reviews
The supplier says waterproof. The reviews on the same item sold under three other brands say it leaks at the seam in month four. Both documents are sitting in the same folder, and putting them next to each other is the cheapest check available before you commit to a first order and a container of it.
Draft the store while the decision is still open
Titles, descriptions, the FAQ, the returns wording and a first ad angle can all be written before you commit, and doing it early makes the decision better rather than premature, because you find out what you would actually have to claim. Drafting costs a model call and touches nothing live.
The numbers nobody actually has
Most disappointment with AI product research comes from expecting data it was never going to hold.
No search volume, no sales estimates
Helium 10, Jungle Scout and Keepa pay for marketplace data you cannot get any other way, and none of it is in here. If the question is how many people searched a keyword last month, that is a subscription. If it is whether the numbers you do have justify the category and what the listing would then need to say, that is this.
Displayed sales and BSR are indicators, not facts
Platform-shown sales counts, best-seller ranks, heat scores, review totals and third-party estimates all carry error bars, and where each figure came from is stated next to it so you can weigh it yourself rather than inheriting someone else's confidence.
It does not place the order
Final sourcing, stocking, ad spend and pricing decisions stay yours by design, because they depend on cash position, supplier terms and how much risk you can stomach, none of which is in any file you gave it. In higher-risk categories such as food, supplements, cosmetics, children's products, medical devices, pet food and electricals, the evidence bar rises and human review is required rather than suggested.
Before you point it at a niche
Does it scrape supplier and marketplace sites?
Not by default and never past a barrier. Logins, captchas, anti-bot measures, permissions and paid API limits are not worked around, and automatic collection needs you to confirm the authorisation, the login state, the platform terms and the rate. Most research runs perfectly well on exports and saved pages, which is what people actually do.
Is this the Claude Code dropshipping setup people write about?
Same pieces, with the terminal part already done. Those guides describe a rules file the agent always reads, skills as task playbooks, subagents researching in parallel, and a confirmation step before anything touches live store data. All four are parts of the app here, and Claude Code itself can be driven from inside if that is what you are used to.
Can it build the Shopify store as well?
It drafts everything that goes into one and it can talk to Shopify Admin through a connector, but that connector is a native integration you set up with your own Shopify application and API scopes, not a one-click link. What it will not do is publish, change prices or spend money without you.
What does one research pass cost?
The app is free and MIT-licensed; the variable is model usage. Twenty competitor pages is a large context read, so before you queue thirty products it is worth knowing whether you are paying your own provider directly or going through an Orkas plan. Neither is priced per product or per report.
Give it the folder. Keep the decision.
Free, macOS and Windows. Your model key or an Orkas plan.