Cardloader

Card recognition software that identifies, grades and prices cards in bulk

Everything Cardloader does, and what each engine can and cannot do. Free local AI for sorting, the Ximilar API when you need exact matches, prices and grades.

Local AI

Offline card recognition on your own computer

Cardloader includes an on-device engine that runs vision language models through llama.cpp: LFM2.5-VL 3B by Liquid AI (the default, about three seconds per card on a recent Mac; the smaller 1.6B and 450M models are faster) and Qwen3.5 4B and 9B by Alibaba’s Qwen team, which read more carefully and take about 7 and 14 seconds per card on a MacBook Air M4.

It uses Metal on Macs, and on Windows and Linux any NVIDIA, AMD or Intel graphics card through Vulkan, which comes with the graphics driver; without one it runs on the CPU. Settings → Local models → Run on picks the graphics card, and a model larger than its memory runs partly on the CPU.

The LFM models take two quick looks at each card: the whole card for its name, year and game, and an enlarged strip of the bottom edge for the collector number and set code. Qwen3.5 reads the whole card in one look. Sports cards print their number, year and set on the back, so in a session of sport cards every model reads the front and the back by default (other sessions can turn it on). Answers are held to a fixed JSON shape, so they always parse into clean columns.

  • Free and unlimited: no API key, no account, no per-card fees.
  • Private: images are read on your machine and never uploaded.
  • One-time download: about 2.3 GB from Hugging Face for the default model (330 MB to 6.6 GB for the others) after you accept the model’s license; it can be paused and resumed.
  • Sports card numbers from the back: on 100 sports cards photographed front and back, Qwen3.5 9B read 88 card numbers right, Qwen3.5 4B 79 and LFM2.5-VL 450M 33. From the front alone, none of them could.
  • Your own model: Cardloader is open source, so a build of your own can run any other vision model from Hugging Face, however large, on your graphics card for free. It takes one entry in the model list, which a script writes for you. Tried with Google’s Gemma 3 4B, it read 19 names and 18 card numbers right on 20 Magic cards.
  • Honest limits: it cannot tell reprints and variants apart, and gives no match score, alternative candidates or prices. Its grades and conditions are rough estimates, not a grader’s opinion. Rows show a Local tag instead of a confidence tier.

The LFM2.5-VL models are licensed under the LFM Open License v1.0, which limits commercial use above a revenue threshold. Qwen3.5 is under the Apache 2.0 license, which allows commercial use, like Cardloader itself.

Compare with local model dialog showing how often the on-device model agrees with the Ximilar API
Ximilar API

AI card identification with exact matches and prices

For the most accurate results, a session can use the Ximilar Collectibles Recognition API. In Auto mode, each card is first classified and then sent to the right endpoint: TCG ID, Sport ID, Comics ID or Slab ID. Cards that lie sideways or upside down are turned upright before they are matched.

  • Confidence tiers: every match is marked high, medium or low, so you review only the uncertain cards.
  • Alternative candidates: if the best match is wrong, pick another one with one click.
  • Price statistics: a representative USD price with links to marketplaces such as TCGplayer and eBay.
  • Pay per use: one API call per recognition (two in Auto mode). Needs a Ximilar Business plan or trial; see Ximilar pricing.
Card detail for a Shohei Ohtani Topps Chrome baseball card with market price, grade and corner, edge and surface scores
Grading

Card grading, condition and centering

With the Ximilar Card Grading API, Cardloader can pre-grade every card in a batch before you decide what to send to PSA, BGS or CGC. Grading overlays are saved with the session so you can see exactly what was measured.

  • Full grade: corners, edges, surface, centering and a final grade, using both sides of the card.
  • Condition: a fast condition check on the eBay, TCGplayer, Cardmarket or Ximilar scale.
  • Centering only: a centering score from both sides, for a quick sort before grading.
  • Local estimate: the local model can also estimate a grade or a condition on the TCGplayer, Cardmarket or eBay scale, for free and offline. It is a rough sort, not a grader’s opinion.
Card types

TCG, sports cards, comics and graded slabs

One scanner app for the whole collection. These are the categories the Ximilar API recognizes; the local model reads the printed text of any card.

Pokémon card scanner for PC and Mac

Identify Pokémon cards in bulk, including set and collector number, from a folder of scans or straight from the scanner.

Magic: The Gathering, Yu-Gi-Oh!, One Piece and Lorcana

The same flow covers the major trading card games, and more are added to the Ximilar API over time.

Sports card scanner software

Baseball, basketball, football, soccer and hockey cards, identified by name, set, year and card number.

Comics and graded slabs

Comic books are identified with Comics ID, and graded slab labels from PSA and other graders are read with Slab ID.

Comics

Comic book scanner: comics, magazines and manga

Cardloader scans comics the way it scans cards. Choose Comics in the setup, or as a session’s Identify as, and every scan or photo of a front cover becomes a row with the series, the issue number, the publisher and the date. Only the front cover is read: the back of a comic says nothing about which issue it is.

With the Ximilar API, Comics ID matches the cover against a database of comic books, comic magazines and manga. It returns the series, the issue number and its title, the publisher and the cover date, with links to marketplaces and, if the session asks for them, price statistics.

The local model reads what the cover prints: the title of the series and the issue number, plus the publisher and the year when they are there. Qwen3.5 reads the whole cover in one look. The LFM models take two: the whole cover for the title, then its enlarged top edge for the issue number, where older comics print it next to the date and the price.

  • Measured: on scans of 11 covers from the 1940s, Qwen3.5 4B and 9B each read 9 issue numbers right and LFM2.5-VL 450M read 4. Modern covers, which often print the number in the barcode box, were not measured.
  • Same table, same export: the series is the row’s name, the issue its number and the publisher its set, in the app and in the CSV.
  • Not yet: comics cannot be listed on eBay from Cardloader.
Sources

Bulk card scanning: scanner, folder or watch folder

Cardloader takes cards from wherever they are. Original images are never modified or moved.

  • Document scanner: drives a RICOH/Fujitsu fi-8170 (or any TWAIN 2.x scanner with a feeder on Windows, any SANE scanner on macOS and Linux) in ADF duplex batches: 600 dpi when you grade, 300 dpi when you only identify. Cards fed upside down or back side up are turned and paired the right way.
  • Folder: analyse JPG, PNG, WebP, TIFF or BMP images from a folder of scans or phone photos.
  • Watch folder: analyse images as another app saves them, for example PaperStream Capture or ScanSnap Home.
  • CSV file: import an existing spreadsheet; rows with an image path are analysed, rows with a card name come in as identified.
New session dialog in Cardloader: folder, watch folder, scanner or CSV source, identification engine and grading options
Front and back

Automatic front and back pairing

Choose one image per card or front and back pairs. Files are paired by name (front/back, recto/verso, f/b, a/b), or in order when they have no such suffix. Duplex scans from the scanner are paired for you.

Home

A quick setup and a home screen

On first launch, five questions set up Cardloader for how you collect: what you scan, which game, whether you grade, check condition or want prices, how the cards come in, and whether the Ximilar API or the local model identifies them. The answers become the defaults of every new session, and you can run the setup again at any time.

Home then shows the cards analysed in the last 7 days, a year of scanning activity, your last sessions and those defaults, with a button to start the next session.

Cardloader home screen with cards analysed in the last 7 days, a year of scanning activity, the last sessions and the defaults for new sessions
Sessions

Sessions you can pause, resume and fix

Every folder analysis or scanner run is a session stored on your computer.

  • Results fill in live; click a card to see its images, match, price and grade.
  • Pause, stop and resume, even after restarting the app. Retry failed cards one by one.
  • Run again picks up new files added to the folder since the last run.
  • Up to 3 cards analysed in parallel for fast bulk scanning.
  • Search, filter, rename, duplicate and bulk-delete sessions.
Selling

List cards on eBay, export for TCGplayer, Whatnot and Shopify

When a session is done, its identified cards can go straight to a marketplace, at Ximilar’s price or yours.

  • eBay: each card becomes a draft listing with its pictures, title, item specifics and condition; graded slabs keep their grader, grade and certificate number. Check the drafts, then publish them, and the session shows which cards are listed and which sold. It uses your own eBay developer keys, and listing fees are eBay’s.
  • TCGplayer: a list of the ungraded trading cards in TCGplayer’s terms, with each card’s TCGplayer product link, to add in Seller Portal.
  • Whatnot: the bulk listing CSV, for auctions or Buy it Now.
  • Shopify: a product CSV with condition variants, or an inventory update.
  • Pictures: for Whatnot and Shopify, the front and back of each card can be uploaded to ImgBB with your own key, so the import finds them.
AI agents

Run Cardloader from Claude and other AI assistants (MCP)

Cardloader has a built-in MCP server, so an AI assistant such as Claude Desktop, Claude Code or Cursor can work with it for you. Ask in plain words, for example “run a new session on my Pokémon folder, English cards, fronts only, and store it in Ximilar” or “review the latest session and accept the cards that match their pictures”.

The assistant sees each card’s match, the other candidates and its front and back pictures, so it can check a match the way you would. Everything it does shows up in the app as it happens.

  • You decide what it may do: reading is always allowed; running and reviewing sessions, deleting, marketplace listings and changing defaults each need their own switch in Settings.
  • AI reviews stay visible: cards an assistant accepted or corrected carry an AI mark and its note until you confirm them.
  • On your computer only: the server listens on 127.0.0.1, needs a token and runs while Cardloader is open. It never reads or changes your API keys.
  • Ready-made tasks: “Review a session” and “New session from a folder” appear as prompts in assistants that support them.

It works with AI apps that run on your computer. Assistants that only connect to servers on the internet, such as ChatGPT connectors, can’t reach it. See how to connect an assistant.

Export

CSV export and import for your inventory

Export any session to a UTF-8 CSV that opens in Excel, Google Sheets or your store’s import tool. Columns include name, set, set code, card number, series, year, category, confidence, price in USD, grade, condition and the image paths. An exported file imports back exactly as it was.

Cloud sync

Optional sync to Ximilar Cards Cloud

Mirror sessions live to Ximilar Cards Cloud, with each card’s image, match, prices and grading, and browse what is stored there from the Cards Cloud tab. Corrections you make in Cardloader are sent along. Sync is off unless you turn it on.

Open source

Open source: adapt it to any collector use case

Cardloader is released under the Apache-2.0 license. Read every line, run it anywhere, and fork it for whatever you collect: coins, stamps, sealed product or your own card database.

It is built with Tauri 2 (Rust) and React with TypeScript, stores sessions in SQLite and keeps your API key in the system keychain. Sources, engines and the CSV export each live in their own module, with tests that include an end-to-end run against a mock API.

Missing a feature? Add it yourself, or ask an AI coding assistant such as Claude Code to do it: the repository includes a guide for AI agents with the code layout, the commands to run and the rules a change has to follow.

Start scanning your cards for free

Download Cardloader for Windows, macOS or Linux, point it at a folder or your scanner, and identify your first batch in minutes. No subscription and no account needed.