100% offline · on-device AI
AI File Renamer reads every document and photo on your own machine — the text, the document type, the people, the dates — and proposes a file name you would actually have typed yourself. Nothing is uploaded. Nothing is renamed until you say so.
Version 1.2.0 · Windows 10 & 11, 64-bit · installs from the Microsoft Store, no
administrator rights.
The Store listing is in preparation — this button goes live with it.
Rename — 3 proposals ready for review
Illustrative example with invented names. Proposals are editable, and nothing is written to disk until you press Apply.
The problem
AI File Renamer
Scan_0001.pdf is a folder you can’t searchA timestamp tells you when you pressed the button — never which certificate, which contract, or whose passport. Finding one document means opening a dozen.
AI File Renamer opens each file, extracts the text (OCR, or the PDF’s own text layer when it has one), works out what the document is, who it names and which date actually matters — the issue date, not the birth date or the expiry date.
The result is a name in the shape you would have written by hand — document type, person, place, date — plus searchable metadata written inside the file itself.
Features
Passports, certificates, medical records and family photos are exactly the files you would never paste into a web service. So none of this needs one.
Recognition, classification and naming all run locally. Your files never leave the machine, and the app works with the network cable pulled out — the only traffic it ever makes is downloading the recognition models you choose to install.
Latin, Cyrillic, Greek, Arabic, Devanagari, Chinese, Japanese, Korean, Thai, Tamil and Telugu recognition packs, installed one at a time from Settings. The script is detected automatically — adding a language is a download, never an app update.
Document type, person, place and date are read off the page and composed into a name. The same facts are embedded in the file — PDF Info/XMP, JPEG EXIF — so they survive copies, moves and cloud sync, and Windows Search can index them.
Every proposal is listed with its confidence before anything changes. Edit a name, untick a row, open the file. Applied batches are journaled, so a single Undo puts every file back the way it was.
Finds byte-identical copies and visually similar re-compressions across an entire tree —
the (1), (2) sediment of years of syncing. Strictly a report:
it groups them for you and never deletes anything.
The optional local language-model tiers — which sharpen difficult names and describe photos — run about three times faster on a Vulkan-capable GPU, and fall back to the CPU automatically when there isn’t one.
How it works
Point the app at a folder. PDFs (including multi-page), JPEG, PNG and TIFF are analyzed; anything else is skipped and reported, never touched. Large archives process incrementally, so a second run only looks at what changed.
Each file arrives with a proposed name, a scan-or-photo classification and a confidence score. Correct anything you disagree with, uncheck what you want left alone. Nothing has been written to disk yet.
Apply embeds the metadata and renames the accepted files through a journal. Change your mind and Undo restores the original names for the whole batch.
Privacy
This is not a policy decision that could be reversed in an update — it is how the app is built. The recognition models run inside the application on your own hardware, and there is no server on the other end to send anything to.
Screenshots
FAQ
Yes. All analysis runs on your machine. The one time the app uses the network is when you click Download to install recognition models (roughly 34 MB for the required set, more for optional languages and the local language models). After that you can disconnect entirely and everything still works.
Recognition packs cover 20+ languages across Latin, Cyrillic, Greek, Arabic, Devanagari, Chinese, Japanese, Korean, Thai, Tamil and Telugu scripts; you install only the ones you scan. The script is detected automatically, and a page it cannot read is reported as unrecognized rather than named from garbage. Hebrew is not supported today — no suitable recognition model exists upstream.
No. Files are renamed in place, and metadata is written without re-compressing an image or re-rendering a PDF. Duplicate detection only produces a report — it never deletes. Undo restores the original names; the metadata it embedded is additive and stays, which is deliberate.
It will be a paid app with a free trial, sold through the Microsoft Store. The trial and the purchase are handled entirely by Windows — there is no licence key to enter, no account to create, and the app itself never contacts us.
Windows 10 or 11, 64-bit. The installer is per-user, so no administrator rights are required. Budget about 250 MB for the app plus the models you install. The OCR engine needs the Microsoft Visual C++ x64 redistributable, which is already present on most systems.
The engine is cross-platform by design and the shared core is built to be reusable on Android and iOS later, but Windows is the only packaged, tested build today. There is no release date for the others.