What's in the Box

A storage register you run from your phone. Photograph an object as it goes into a storage location — a box, a shelf, a cupboard, whatever holds it — and the app identifies it, files it against that location, and lets you search months later for where a thing ended up. The app's own name is the one place "box" still literally means a box; everywhere in the UI it's a "location" (LOCATION-0001 on the printed labels), on purpose, so it fits a shelf or a cupboard just as well.

Six static files. No server, no build step, no framework.

How it fits together

Phone (Safari on iPhone, Chrome on Android; installed to home screen)
  |
  |-- static files ......... GitHub Pages or Cloudflare Pages
  |-- one call per photo ... api.anthropic.com
  |-- book lookups .......... openlibrary.org (books only, no key needed)
  |-- sync, if connected .... login.microsoftonline.com, graph.microsoft.com
  |-- catalogue + photos ... IndexedDB on the device

Everything this app talks to, and why, is also listed on the Setup tab under "Connected services."

Nothing of yours listens for inbound connections. The phone makes outbound calls only, the same way any app on it does. The static host holds no data and runs no code of yours.

Files

File	What it is
index.html	The whole app: UI, database, camera, API calls, label printing
qrcode.js	QR encoder (MIT, Kazuhiko Arase), vendored so labels work offline
manifest.json	Makes it installable to the home screen
sw.js	Service worker; caches the app shell so it opens without signal
icon-192.png, icon-512.png	Home screen icons
CLAUDE.md	Project context for Claude Code

Hosting

The site has to be served over HTTPS, otherwise the browser refuses to register the service worker and the app cannot be installed.

Private repo, free: Cloudflare Pages. It connects to a private GitHub repo on the free tier and redeploys on every push. This is the recommended route.

Public repo, free: GitHub Pages. Settings, Pages, deploy from main at root. Nothing sensitive is in the repo, so this is a legitimate option.

Private repo on GitHub Pages requires a paid plan, and even then the published site is still publicly reachable. Only an organization on Enterprise Cloud can serve Pages privately.

In every case the site is public; only the source can be private. That is harmless here, because the page is an empty shell and all data stays on the phone. If you want the URL itself gated, Cloudflare Access adds a login free.

Setting it up

1. Push the repo, connect it to Cloudflare Pages, open the URL on the phone.
2. iPhone: Safari, Share, then Add to Home Screen — Chrome on iOS cannot install PWAs. Android: Chrome will offer to install it directly.
3. Setup tab: paste an Anthropic API key from console.anthropic.com and save. Set a monthly spend limit in the Console while you are there.
4. Locations tab: pick or create the house you're packing at (top of the tab) — a location always belongs to whichever house is selected there. Then create a location (a box, a shelf, a cupboard — whatever it actually is), pick a colour, print its label, tape it on. The "copies per label" choice controls how many identical copies print on one sheet — enough to tape on more than one side, or more than one shelf edge — sized to use the page well whether you pick 2, 4, or 6 per sheet. Sheets print at A4 with zero page margin (the printable area is set by the label sheet's own layout, not your printer's default margin) — if your printer's dialog offers a margin or scale option, leave it at Default margins / Actual size, not "Fit to page" or "Shrink to fit", or the label sizes will be off. If a label's colour bar or the scale card's grey/black bar prints blank, look for a "Background graphics" toggle in the print dialog and turn it on — the app already asks for this in its CSS, but a few browsers still require it set manually too.
5. Capture tab: one object per photo as it goes in — only locations at the currently-selected house show up here, since you can't be physically standing at the other one.
6. Find tab: search by name, material, or text printed on the object — this searches across every house, each result labelled which one it's in.

Book lookup

Anything identified as a book gets an automatic, free lookup against Open Library (title/author, no account or key needed) — shown under the item as supplementary info, never overwriting what Claude itself identified. Open Library's data isn't always right (a real ISBN looked up while building this returned a completely different, mismatched book), so treat it as a second opinion to glance at, not a fact.

If the automatic lookup doesn't find a match, a "Photograph copyright page" button appears on that item — take a photo of the page inside the book with the ISBN and publisher details (more reliable than the cover), and it looks up that exact edition instead of guessing from the title.

Anything that looks unusual — a book with very few editions on Open Library, or Claude noticing a first-edition mark, signature, or similar while identifying any object — shows as "Worth a look." That's a nudge to look closer yourself, not a valuation; nothing here can tell you what something is actually worth.

Coin lookup

Anything identified as a coin can be looked up against Numista's catalogue — but not automatically off a single photo. A coin's date or mint mark can be on either face, so a search run before both sides are even in wouldn't necessarily find the right entry, and searching twice (once per photo) would just waste your Numista quota on a result about to be replaced. So: taking one photo identifies the object (title, material, etc.) same as anything else, with no coin match yet. Add a free Numista API key under Setup, "Coin lookup", to enable matching at all — no key, no lookup, ever.

Once a photo is identified as a coin, the Capture tab itself offers "+ Add another photo" — flip it over and photograph the other face there and then, no need to switch tabs. (The same prompt appears for a book too, useful for its copyright page; it's hidden for everything else, since most objects only ever need the one photo.) The same button also appears on any coin item afterwards, on the Find tab. Either way, Claude re-identifies the item using both photos together, and about a second and a half later the Numista search runs on its own using that fuller picture. If you only ever take one photo, that's fine too — nothing forces a second — but it means no automatic match; tap "Retry coin lookup" (Find tab, appears on any coin without one) whenever you want a search run on what you've got.

A text search has no way to guarantee its top result is the right coin — confirmed a real case where it wasn't, during testing, a wrong match on the reverse of a 20-cent-euro coin. So the automatic search and the quick "Retry coin lookup" button both take the top result on trust (the same tradeoff Open Library's book lookup already makes), but the item detail view's own search (tap any coin, "Search again") shows up to six candidates to choose from instead of picking one for you — pick the right one directly, rather than trusting a guess.

Once matched, the coin's N# is a clickable link back to the real Numista page, and Numista's own photo of that catalogue entry shows right there so a match can be visually checked at a glance. Opening the item's detail view also loads the fuller picture from Numista — the denomination, composition, and each side's own recorded lettering and design description — to compare against what's actually on the coin.

If the text search doesn't find a match at all, it just stays unmatched — there's no automatic image fallback, because Numista's image-search endpoint needs their paid plan (€100/month minimum). Setup has a "Run image search on unmatched coins" button instead: it shows you the real estimated cost first, and is meant to be run as an occasional batch once a backlog of unmatched coins has built up, not per-coin. Numista waives the €100 monthly minimum for the first calendar month a paid plan is active, so the intended pattern is: activate the plan, run the batch once, then cancel or downgrade before the next billing month.

Sending a second photo to Claude roughly doubles that item's identification cost, same as any extra image in the request; the Numista search itself, automatic or manual, is free either way.

Item detail

Tap any item on the Find tab (not its photo-strip's delete button) to open a full detail view: every field Claude filled in, both photos if there are two, and — for a coin — three editable search fields (the text sent to Numista, a year, and a size in mm, the last two pre-filled from whatever's already known) with a "Search again" button. Searching there shows up to six candidates to pick from, not just one guess, so a wrong top match (a real 20-cent-euro-coin test during development landed on the wrong design entirely) can actually be corrected — free, no new photo, no re-identification needed. Nothing else on this screen is editable yet; deleting and re-capturing is still the way to fix a wrong category, material, or title.

Photo scale card

Setup, "Photo scale card" — prints two L-shaped corner rulers side by side, one labelled OBVERSE (front/"heads") and one REVERSE (back/"tails"), each framing a dotted circle to place a coin in. A corner ruler reads width and height from one photo, unlike a single ruler laid alongside an object which only gives one axis, and the alternating black/grey bar above each ruler's numbers gives a size at a glance without reading them. Photograph the coin in the correctly labelled patch for each side — the printed label itself is what tells Claude which side it's looking at, since it's right there in the photo. Print it once and reuse it. Same "Actual size, not Fit to page" rule as the labels applies.

Syncing two phones (OneDrive)

Setup is a one-time thing, already done for this deployment: an Azure app registration (personal Microsoft accounts, SPA redirect URI, the narrow Files.ReadWrite.AppFolder Graph permission — not the broader Files.ReadWrite) with its client ID wired into MS_CLIENT_ID in index.html. No cost, no client secret.

On each phone, under Setup:

1. Set a sync passphrase — the same one on both phones. It encrypts everything before it ever reaches OneDrive, so OneDrive only ever holds unreadable bytes. Put it in a password manager: if it's lost, whatever synced under it is unrecoverable, by design.
2. Connect OneDrive — sign into the same dedicated Microsoft account on both phones (not your own personal one). That's what makes the narrow Files.ReadWrite.AppFolder scope possible — the app can only ever reach its own folder, and there's nothing else in that account to reach anyway. No folder-sharing step needed; both phones are automatically looking at the same app folder.
3. Sync now whenever you want the two catalogues to merge. It also fetches on its own each time the app is opened with a connection.

Resetting the catalogue

Setup, "Danger zone", "Reset all data" — wipes every house, location, and item on this phone. It does not touch your saved API keys, sync passphrase, or OneDrive connection, so you won't need to re-enter those afterwards. Before anything is deleted it downloads a full backup (catalogue + photos) automatically, and asks you to confirm twice — the second time by typing DELETE. If the other phone is synced, a reset here is recoverable by syncing again; if it isn't, the downloaded backup is the only copy, so keep it somewhere safe.

Local development

python3 -m http.server 8000    # http://localhost:8000

localhost counts as a secure context, so the service worker and install prompt both work with no certificate. After changing sw.js, bump the CACHE constant or the old shell keeps being served.

Cost

Photos are resized to 1024 px before sending, landing around 1,400 input tokens each. On Haiku 4.5 that is roughly one to two euros per thousand objects. The rarity/"worth a look" check rides along on that same call (a few dozen extra tokens, not a second call), so it doesn't meaningfully change this. Book lookups and the automatic coin text lookup are both free and don't touch the Anthropic bill at all. The optional coin image search is the one paid extra in this app, entirely separate from the Anthropic bill and never run without a confirmation showing the real estimated cost first — see "Coin lookup" above.

Known limits of this prototype

• Data lives in this browser's IndexedDB. iOS can evict it after weeks of disuse. Export from the Setup tab regularly even with sync configured.
• Sync doesn't propagate deletions yet — deleting an item on one phone can have it reappear from the other phone's next sync. Documented, scoped limitation in CLAUDE.md, not a bug waiting to be found.
• Labels print but are not scanned back in; you pick the box from a list.
• Item fields are not editable after identification, only deletable.
• Photos are stored as data URLs, about a third larger than raw blobs. This matters more the bigger the catalogue gets — see the roadmap note in CLAUDE.md if you're aiming for a large one.
• Search is substring matching over the stored fields. Fine to a few thousand items, and the point at which that stops being true is the point to add embeddings.

Background

Earlier designs, and why they were dropped:

• FastAPI on a Mac, phone as a client over Tailscale. Worked, but meant a machine to keep running, patched and certificated, holding the only copy of the database. Too much to own for a tool used while carrying boxes.
• A local vision model on a GPU. Same objection, plus the phone then only works when that machine is awake.
• Native Swift with SwiftData and CloudKit. Genuinely better: real sync, no data eviction, free on-device OCR and image feature prints through the Vision framework, background uploads that finish when the app is closed. The cost is Swift, Xcode, and €99/year. Still the right destination if this proves worth keeping.