The scan inbox

Everything you scan lands here first. Each row says what it needs in one word, and tapping that word is the whole job.

1. Add files the row where it says, and Install & add installs that table first when it is a ready-made one you have not set up yet.
2. +1 more means you already have this. Say which of three things happened and the count goes up the right way.
3. Review in amber means something needs a look. The row asks one question with the answers as taps.

A session's File N button does the same for every row whose own button would file, and reads Install & file N when filing will install something.

What a row shows

• Its picture, its name, and where it is going. The table pill picks any table you have (choosing one never files), and the button beside it is the row's state. When a table set up after the scan fits better, it is offered as a chip beside the pair, "Groceries?", and a tap moves the row without filing.
• Fields as chips, sized to their values. A dotted value is Cobblr's guess, a solid one is yours. Tap to edit. The ones nothing filled in are offered underneath by name.
• Where a value came from. Every field a scan fills records its source: the code, the photo, the receipt, or the AI's own knowledge. A row lists the ones the AI supplied rather than read, and a field the item itself contradicts (an author printed on the cover that is not the author the AI gave) holds the row for review with both values as taps, whatever the confidence.
• Names read with a first capital ("corn starch" shows as "Corn starch") and the stored name is untouched. A name you typed is shown exactly as you typed it, and a re-run of the AI never writes over it: when its read differs, the card says AI read "…" with Use it beside your name.
• A title in another language keeps its other forms. A book filed as Желтый туман is found by searching "Yellow Fog", and how it reads everywhere (original with the translation in brackets, the translation first, or the transliteration) is your Titles in another language setting under Your account → Appearance. Renaming still edits the stored name.
• A name only the web could find is a tap, not a retype. When no catalog knows a barcode (an earbuds box, a cable, a pack of batteries) and one web page names it, the card asks Is this it? with the name as a tap, marked as found on the web and not confirmed. Typing your own stays beside it, and the name is never asserted for you.
• A field you are typing in is yours. The inbox refreshes while the AI works, and a box you are in keeps what you typed. A value that changed underneath is offered as updated underneath with Use theirs and Keep mine, only when it actually differs.

On a phone

The inbox is a list of compact cards, and an item opens as its own screen.

• The card shows the picture, the name with a Details link, a few of the table's fields, then one row with the quantity (− ×1 +), a camera button for a better photo, Re-run AI, and a ⋯ sheet with the rest. The last row is the filing pair.
• The item screen has a pinned header (All items to go back, where you are in the queue, previous and next), the name with its barcode (tap to correct the code), your photo beside the catalog picture, the table and every one of its fields as a form, What the AI found folded until you open it, Correct it with a research hint and Re-run AI, and a pinned footer: Discard, Cancel, Confirm.
• A dropdown grows from the form. Every choice field ends with New category… (the field's own word). Pick it, type the name, and Enter adds it to the field for everyone and selects it here.
• An empty picture slot is a tap that opens your camera roll, and the picture lands where the slot says.
• A value you type stays on the item, filed or not. Move on, close the screen, come back at a desk and it is there, in what Add, File N and the assistant write. Clearing the field puts the AI's value back.
• The top of the queue is four short lines: the put-away strip, a series offer, and each session's header (who it is, the count, the location pin, File N, and the open arrow).

At a desk

A card is as tall as what it has to say. The camera, ⋯, Re-run AI, discard and the expand chevron sit in one row at the top right, and the title edits in place with a tap (Enter saves, Escape reverts).

A row that needs a look asks its one question where the AI note used to be: Check weight: 1 - Fingering (the name says "4 Ply") or 4 - Aran (the AI's read). Tapping an answer sets the field, and when it was the last open question the row offers Add in the same tap. The full note is still on the open card under Source data.

Filing a session

• One session files as a batch. Once identification settles, File N files the ready rows. Rows that need a look are not swept in.
• The location dialog names where things are about to go, and carries Review the plan first, which opens the sorting plan for that session. The put-away strip at the top of the page is the same door for every session.
• A mixed batch asks per table. Agree about the teas, disagree about the spices, and each group goes where you said. Closing the dialog files nothing.
• File everything sorts the whole inbox into three piles: added to what you have, filed as new, and left for you with a reason. It writes nothing until you confirm. See putting things away.
• A scan that names one of your own lists goes there. Green tea whose category is "Teas" files into your Teas list, not into the catch-all. The catch-all shows a strip for anything already filed the old way ("2 items → Teas"). One tap selects them and opens Move to…, and the move keeps history, tags and printed labels.

Things you already have

Scan something you track and the row reads +1 more instead of making a quiet duplicate. Its line names the record with its count and where it lives ("You already have Green tea ×1 · Pantry"), and offers the three answers that matter:

• Replaced the one that ran out. The count stays what you scanned, and the last one is recorded as used, which is how Cobblr learns how long a box lasts.
• +1, still had some. It goes on top.
• Old one went bad. Recorded as waste, never as consumption.

Any of these dates the new stock properly: a fresh lot from today, good until today plus its shelf life, with its bought-on date. Sending the scan back removes that lot again, or takes back what is left if you have used some since, and takes back the purchase it recorded, so nothing Cobblr learned from it survives an undo.

Two more offers appear when they apply. When the record it matched is filed somewhere else (Tea, but filed in Inventory), the banner offers Move it to Tea with the destination already chosen. When the matched record has no picture and the scan just found one, it offers Use this scan's picture.

A one-of-a-kind thing (a vehicle, a tool) is asked "is this the same one?", and a merge fills only blank fields, so nothing you entered is overwritten.

Two rows that are the same thing

One thing can end up as two rows: a shop parsed on Monday calls it "Roma Tomatoes" and one on Friday calls it "Tomatoes Roma". New scans no longer add to the pile, since a match is counted onto what you have. For the pairs already in your tables, open the page's ⋯ menu and choose Same thing twice. Each pair shows both records, how many are on hand and where they live, and the words they share. Press Keep this one on the survivor: the quantities are added together, anything the kept record is missing is taken from the other, a date takes the earlier of the two, and the other record is removed. The kept record keeps its id, its links and its history. If a merge cannot finish, nothing changes.

Pictures

• Your photo beside the catalog picture. Tap either to see it full size, where Rotate, Crop for catalog and Use as catalog live. The tile that is the catalog picture carries the accent outline.
• Crop for catalog opens your photo full size. Drag a box across the part you want and press Use crop as catalog image. The crop is cut from your full-resolution original, so cropping again starts from the original. Revert steps back through it like any other pick.
• Paste a picture, or its link. Paste an image address into the image search box and Cobblr fetches that exact picture. The button reads Use instead of Search as soon as what you typed is a link. Ctrl+V with a picture on your clipboard, anywhere in the inbox that shows photos, makes it the item's catalog photo.
• A picture you pick is remembered. The next scan of the same product gets it rather than another web guess. For a barcoded item that goes through the shared product database. A receipt line has no barcode, so the shop plus the line's name is its identity, and a picture picked for one shop's "Baby Carrots" does not answer for another's. Only pictures from the web are shared, and a photo you took stays in your workspace.
• A row with no picture keeps looking on its own, right away once the web search is answering again, and otherwise every six hours for a week. A picture you chose is never replaced. Fresh produce is looked up in a photo library first, so "roma tomatoes" comes home as tomatoes and never as a tin, and when the web search is rate-limiting the strip says so instead of claiming nothing exists.
• Photos you already took are one tap from the dashboard: the Add photos tile beside the scan shortcut opens the same picker the inbox uses. They land as one new session with identification already running. When Cobblr is installed as an app, long-pressing its icon offers Add photos on Android and desktop.

Tools

The page's ⋯ menu holds the rarer things: Same thing twice, the First-look question switch (owner or admin), Export and Import for moving a whole inbox between workspaces (photos, history, your typed hints, and every session travel), and Show as for the list or gallery.

A row's own menu offers the tools it could actually need first, from what it knows: two things in the picture offers Split into items, a picture that reads as a receipt offers Read as a receipt, a tote by name offers Turn into a bin. The rest wait behind More tools with the reason each is folded. A laser engraver on the bench is not asked whether its box is empty.

• Split into items cuts one photo of several things into separate rows, each cropped and sent to find its own picture. A split child shows its crop and the group shot it came from (badged "group"), with Crop again to cut it differently. Undo split on any piece brings the photo back exactly as it was. A piece you already filed blocks the undo until you send it back, and the message names which one.
• Turn into a bin makes a storage container a location ("Bin N", renameable), keeping the product's photo, barcode and brand as the bin's identity. Create & scan into it makes it the standing file location.
• The camera button on a closed card means "I will photograph this myself". On a phone it opens the camera ready for that item. At a desk it marks the item, the mark survives a re-run, and the dashboard lists marked items under "needs you" with a Photograph button.
• Things scanned together can be offered a shared tag or category, on the session you are looking at, when you ask for it. It costs nothing until you press it and only says things look related when something was found to relate them.

The first-look question

Switch on First-look question in the ⋯ menu and a photo with no barcode gets a fast first read the moment it lands. The capture card asks "Is it X? Yes / No" while you are still holding the thing. Yes settles the name and the full read fills in the rest. No lets you say what it is instead, with the guess ruled out. Ignore it and the full read runs on its own. It is off by default because it is an AI call on every photo, and it is its own job on the AI page so it can run on a cheaper or local model. Splitting a photo is its own job there too, Split a photo into items, with its own line in AI activity.

Typing instead of scanning

What you type into the dashboard's box, or the scanner's, decides what runs. A product code (a UPC, an ISBN, a URL) goes to the catalog like a scanned one. A model, part or serial number ("PB287Q") is looked up on the web, and a name the web agrees on replaces the code, with the product's picture. Plain words ("a box of screws") go straight to routing, which then finds a picture for what it chose. A lookup that finds nothing keeps the words you typed.

Does photo identification work if I run it myself?

Yes, with your own key. Barcodes, ISBNs and typed names work with no AI at all. Naming a thing from its photo, reading a receipt, and checking a barcode against the photo it came with need a model, and on your own install that is a personal connection: add a provider under Your account → Connections (a free Google AI Studio tier, Ollama on your own machine, Anthropic, or any OpenAI-compatible endpoint), grant it to the workspace, and every scan uses it. Nothing leaves your install until you do that.

Some deployments come with identification built in, switched on by their operator for particular apps or named workspaces. A workspace with its own connection keeps using that, and any failure falls back to it. Built-in identification has a daily allowance, and past it scans still work and wait for you to fill in details by hand.