AI File Sorter: content-aware renaming and sorting on your own machine

The problem AI File Sorter is aimed at

Most file organizers work from rules you write in advance: extension, date, size, a regex on the name. That breaks down the moment names stop carrying information. IMG_2048.jpg tells you nothing about what is in the picture, and a scanned invoice named scan_0031.pdf tells you nothing about the vendor. AI File Sorter's premise is that the file itself carries the signal, in pixels or in text, and a model can turn that signal into a category and a readable name.

The README frames the audience around three kinds of mess: a Downloads folder, an external drive, and NAS storage. Those are the places where files from many sources accumulate with no shared naming convention. The app also targets media libraries, where the metadata inside supported audio and video files can be turned into a name such as 2024_artist_album_title.mp3. If you already have a working convention and a script that enforces it, this tool is solving a problem you do not have.

How the analysis pipeline actually works

The documented flow has four steps. You point the app at a folder or drive. Files are analyzed with the selected local or remote model, and for images the content itself is read. Category and rename suggestions are generated. You review and adjust before anything is changed.

The interesting part is what feeds the suggestion besides the model. The README lists optional whitelists, recent similar results, and your own approved review decisions as inputs. That is a feedback loop: categories you accept become context for later files, which is how consistency is supposed to survive across a large batch. The app also groups files using names, file types, folder context and past sorting results, so two files that look nothing alike can still land in the same category if the surrounding folder and history point that way.

There is a separate cache for categorization and learned behavior, which the README covers as its own topic. That is the mechanism behind the consistency claim, and it is also the thing to watch: a cache built from early mistakes will keep feeding those mistakes back. The README offers category whitelists and a category language selection as the controls here, and whitelists are the blunt instrument that works when the learned behavior drifts.

For images, the app uses built-in visual LLM backends, and the README names Vulkan, CUDA and Apple Metal among the platform targets. For documents, a text LLM reads the content. These are two different model paths, and the README documents the required files for the visual path separately, which means image analysis can be unavailable while document analysis still works.

Installing AI File Sorter and a first real run

The README points to packaged builds rather than a source build for normal use. Windows users can get it from the Microsoft Store, and there is a SourceForge download button for the project. The README has separate Installation sections for Linux, macOS and Windows, so check the one matching your platform before assuming a package manager is involved.

Before analyzing anything, run the compatibility check. The README documents a system compatibility check and shows a compatibility benchmark dialog in its screenshots, which is how you find out whether your GPU path is usable.

# From the README's Installation section for your platform:
# Linux, macOS and Windows each have their own steps.
# Windows also offers a Microsoft Store listing.

The README's Safe First Run advice is specific and worth following literally: copy 20 to 50 files from Downloads, screenshots, photos or documents into a temporary folder, run the analysis, and inspect the review table before applying anything. Nothing is moved or renamed until you approve it, and with a local model the files stay on your computer.

If you do apply a run and regret it, the documented reversal is a single menu action.

Edit -> Undo last run

That is the only rollback path the README describes. It covers the last run, not a history of runs.

For remote models, the README documents three options: your own OpenAI API key, your own Gemini API key, and a custom OpenAI-compatible API endpoint. An internet connection is only needed if you choose one of those. If you go the custom endpoint route, you are responsible for whatever that server does with filenames and file content.

Where the design gets in your way

The review step is the feature and the cost. Every batch waits on a human reading a table. On a Downloads folder with a few hundred files that is tolerable. On a NAS share with tens of thousands, it is a multi-session project, and the README does not describe a batch approval mode or a scheduled run.

Rollback is thin. Edit -> Undo last run reverses one run. If you approve run one, then run two, then discover run one misfiled things, the README does not document an undo for run one. The safe pattern is to apply changes in small increments and verify between them, which is slower than the workflow most people want.

Local analysis depends on model files being present. The README has a section on required visual LLM files, which implies the visual path is not self-contained in the installer. If those files are missing, image analysis is the part that stops working.

Finally, the app is a desktop GUI in C++. There is a Testing section and an optional headless live LLM test path, but the README does not present a general-purpose CLI for scripting sorts into an existing pipeline. If your workflow is cron plus shell, this is the wrong shape of tool.

AI File Sorter against a rule-based organizer

The honest alternative is not another AI sorter. It is a rule-based tool such as Hazel on macOS or a small script built on exiftool and file metadata, which is what most people already reach for.

The difference in approach is where the decision comes from. A rule engine decides from properties you can enumerate: extension, creation date, camera model, EXIF tags, filename pattern. It is deterministic, it runs unattended, and when it misfiles something you can read the rule and fix it. AI File Sorter decides from content, which is the only way to categorize a photo of a receipt versus a photo of a lake when neither has useful metadata. The trade is that the decision is a model output, it varies with the model you pick, and it needs a human in the loop to stay trustworthy.

A second practical difference: rule engines are cheap to run on every file every night. AI File Sorter's local path costs GPU or CPU time per file, and its remote path costs API calls. That makes it a periodic cleanup tool rather than a background daemon. If your files arrive in a steady stream and you want them filed on arrival, rules win. If they have already piled up and the names are useless, content analysis is the only thing that can help.

Licence, maintenance and the upgrade cost

AI File Sorter is licensed AGPL-3.0. For an end user running the desktop app, the practical effect is that the source is available and modifications stay under the same licence. If you are considering embedding this in a product or a hosted service, AGPL-3.0 is the licence family that raises network-use questions, and that is a conversation for your own counsel rather than something to settle from a README.

The repository is not archived and the last push was on 2026-09-09, with v1.9.2 released on 2026-08-14. Releases have been frequent through the 1.9.x line, and the project keeps a CHANGELOG.md at the repository root, so upgrade notes are at least in one known place. Upgrading a desktop app that keeps a categorization cache is the part to think about: the README documents the cache as learned behavior, and it does not say whether a version upgrade migrates or resets it. Clearing or rebuilding the cache after a major version bump is the cautious move, and the README's whitelist and category language controls are what you have to steer it back if the learned categories no longer match.

One more cost that is easy to miss: the README points to a TROUBLESHOOTING.md and a TRADEMARKS.md alongside the licence. The trademark file means the name and logo carry their own terms separate from the code licence.

Editorial conclusion

Adopt AI File Sorter if you have a cluttered Downloads folder, an external drive or a NAS share full of inconsistent names, and you want to keep the files on your own hardware while a model proposes categories. Skip it if you need unattended, scheduled sorting of a live directory tree, because the README describes a review-and-confirm workflow and the only reverse step it documents is Edit -> Undo last run. Before committing to a large archive, verify three things on a 20 to 50 file test folder: that your hardware passes the compatibility benchmark, that your preferred local model files are present, and that the suggested categories match how you actually file things.

Frequently asked questions

Official sources

1. hyperfield/ai-file-sorter on GitHub
2. License: AGPL-3.0
3. Project website
4. README
5. Releases