The four jobs behind every listing,
handed to one AI team.
Most guides tell you to assemble your own kit: a chat tool, something to read your seller data, something else to keep it from doing anything rash. Orkas is one desktop app instead. Put your product notes, your Seller Central exports and your photos in a folder, point it there, and it works on them where they are.
Most weeks come down to these four
Decide whether the product is worth doing
Give it a keyword, an ASIN, a category, or the competitor notes you already keep. You get back the price band buyers actually sit in, who is holding the top spots and on how many reviews, a margin sheet with the assumptions written out, and the sourcing and compliance risks. Anything it inferred rather than read is labelled as such.
Write the listing
Title, bullets, description, backend keywords, FAQ, and the A+ module order with a shot list for the photos. It also writes the Shopify, Shopee, AliExpress, Taobao and Pinduoduo versions from the same product, since each one wants a different shape. Before you paste anything in, it flags the superlatives, the invented scarcity and the borrowed brand names that get listings pulled.
Get the photos and the video made
It starts with the reasons someone would buy, then writes the main-image and detail-page copy around them, the prompts for the product shots, and the hook, shot list and closing line for a video ad. You approve the direction before it spends anything on the expensive image calls.
Find out what the bad reviews are actually about
Hand over the review export. Buyer names and order numbers are stripped first, then the complaints are grouped and ranked by how much they cost you, each one tied to something you can change: a bullet, the main image, a support reply, or the product. If the file has no rating column, it says the sentiment is its own reading rather than the buyer's score.
Four things that decide whether you can leave it alone
An assistant you have to supervise line by line is slower than doing it yourself. These are the four reasons you would not have to.
It knows which job it is on
A prompt you paste in is gone when you close the tab, and you have to remember to paste the next one. Here the compliance pass on a listing, the redaction on a review file and the approval before a paid image call all happen because that is what those jobs require, not because you remembered to ask.
- built in
- no reminding
Your files stay where they are
Point it at the folder that already holds the business report, the advertising export, the review file and the product photos. There is no workspace to upload into and keep in sync afterwards. It reads them off the disk and writes its output next to them.
- local folder
- CSV · XLSX
Nothing reaches the live listing without you
Writing, deleting, spending and anything that touches a live store all stop for a confirmation, and the ones you cannot undo get a separate one. Drafting a title and publishing it are two different acts, and only one of them can cost you the buy box.
- confirm first
- undo-proof
The work is handed along inside one conversation
Whoever researched the category passes the price band and the recurring complaints to whoever writes the listing, and the reasons to buy go on to whoever plans the photos. You watch it happen in one thread instead of copying answers between four browser tabs and losing half the context each time.
- one thread
- several roles
Three things sellers hand over first
A category you have not decided on
Hand over whatever you have: a keyword, a handful of competitor ASINs, a quote sheet from a supplier. What comes back is the price band, how many reviews the top sellers are sitting on, a margin sheet you can argue with because the assumptions are written down, and a list of what is still unverified.
- price band
- margin sheet
A listing that stopped selling
Give it the current page, a competitor page and the last three months of reviews. It looks at what the title is spending its characters on, which buying objection the bullets never answer, and which A+ module is doing nothing, then rewrites and runs the compliance pass over the result.
- audit
- rewrite
Three hundred reviews and a 3.9
The rating tells you something is wrong and nothing about what. Export the reviews and the complaints come back grouped and ranked, with the ones a packaging change fixes kept apart from the ones caused by a promise the listing should never have made.
- grouped
- ranked by cost
What sellers ask before installing anything
Does this replace Helium 10 or Jungle Scout?
No. Those tools pay for marketplace data you cannot get anywhere else: search volume, BSR history, competitor sales estimates. Orkas has none of it. What Orkas works on is the data already in your hands or one export away, and what it does with it is the thinking. If the question is how many people search a keyword, that is a data tool. If it is whether these numbers justify the category and what the listing should then say, that is this.
Can it connect to Seller Central, or do I export everything by hand?
Both work. There is a built-in connector for Seller Central that runs through an SP-API application you register yourself, alongside ones for Shopify, eBay, Etsy, Walmart and WooCommerce. Your credentials stay encrypted on your own machine and buyer personal data is never requested. Setting that up takes a developer registration, so plenty of sellers skip it and just drop the reports they already download into the folder, which works the same way.
Can it actually make the product photos, or only describe them?
It makes them. Your own product photo goes in as the reference, so the shape, the label text and the colour come out yours rather than a lookalike, and the models behind it are GPT Image 2, Seedream 5.0 and Nano Banana. Video works the same way. If no provider is reachable it tells you the image was not made, instead of reporting success on something that does not exist.
What does it cost?
The app is free and the source is public under MIT. Model calls are not free for anyone, us included: connect your own OpenAI, Anthropic or Google account and pay that provider at their rates, or take an Orkas plan and pay us instead. Nothing is charged per listing or per report.
Where does my product and review data end up?
The files stay in your folder and are read from disk, not copied into a library somewhere. The conversation is a different matter: whatever the agent quotes out of those files goes to the model provider you picked, which is your own account if you connected one and us if you are on a plan. Review exports carrying buyer names, order numbers or addresses get those stripped before any analysis starts.
I already use Claude Code. Do I have to give that up?
No. Orkas can drive Claude Code or Codex from inside the same conversation, so they become one more pair of hands rather than a separate window. The merchandising work described above is packaged the same way Claude Skills are, which is why anything you have written for that format keeps working here.
Start with the reports you already download
Free and open source, macOS and Windows. Use your own model account or ours.