What Is AI File Sorter?
AI File Sorter is a desktop application that suggests categories, subfolders and cleaner names for the files in a folder you select, and moves or renames them after you confirm. Its README describes it as a cross-platform tool for content-aware file organization and renaming; the source code is open under the AGPL-3.0 licence.
Rather than relying only on fixed rules, the README says it combines the model's suggestions with optional whitelists, recent similar results and your approved review decisions. The website footer credits HF Studio as the publisher.
Get It
AI File Sorter is installed like any desktop program: a Windows installer or portable build, macOS disk images, and Linux packages, all listed on the official download page.
Windows
- How to get it:
- Installer or portable build for Windows 10 or later on the download page
Windows (Store)
- How to get it:
- Microsoft Store; per the download page, no CUDA support, Vulkan acceleration works
macOS
- How to get it:
- Disk images for Apple Silicon (M2 to M5, and M1) and Intel on the download page; not signed with an Apple certificate, so approve it once
Linux
- How to get it:
- Packages for Debian, Ubuntu, Fedora, Rocky and Arch on the download page
Website
- How to get it:
- filesorter.app
Privacy policy
- How to get it:
- No page on the website; a one-sentence statement is linked from the Microsoft Store listing
Source code
- How to get it:
- github.com/hyperfield/ai-file-sorter
Version reviewed: v1.9.2, published on 14 August 2026 per the GitHub release; the Microsoft Store listing shows the same update date. The GitHub release lists no downloadable files, so installers come from the website, the Store and SourceForge. No terminal is needed; building from source requires a C++ toolchain and Qt 6.
Requirements
No RAM or VRAM minimum for the local models is documented in the README, FAQ or website; the figures that exist describe the app and optional GPU use.
Operating system
- What the sources say:
- Windows 10 version 1809 or later per the Microsoft Store listing; macOS and Linux builds on the download page. The README says macOS source builds target Apple Silicon on macOS 15 or later, while the download page also lists M1 and Intel disk images
Memory
- What the sources say:
- The Store listing gives 750 MB minimum and 1 GB recommended for the app itself; no figure is given for running the language models
GPU
- What the sources say:
- Optional. CUDA 12.x for NVIDIA or Vulkan 1.2 or later, plus Metal on Apple hardware; the app falls back to the CPU. The README says 8 GB or more of VRAM gives a smoother experience, while 4 GB often means partial offload or CPU fallback
Image analysis
- What the sources say:
- The FAQ says a discrete GPU with about 6 GB or more of video memory makes image scanning much faster, and that it can be much slower on the CPU
Models
- What the sources say:
- Downloaded from the Select LLM dialog into a storage folder you can change; download sizes are not stated in the pages read
Hardware check
- What the sources say:
- A built-in System compatibility check times a small workload and suggests a local model for your machine
These figures come from the project and the Store listing; PromptQuorum did not measure memory use or speed on any hardware.
How to Get Started
These steps follow the README's "Safe First Run" advice; PromptQuorum has not run them.
- 1Install the app
Why it matters: Download the build for your system from the download page; on macOS, approve the unsigned app once under Privacy & Security. - 2Choose a model
Why it matters: Open Settings, then Select LLM, and download a local model, or enter your own OpenAI or Gemini key or a custom endpoint instead. - 3Try a small copy first
Why it matters: Copy 20 to 50 files into a temporary folder, select it, tick the options you want and click Analyze. - 4Review, then apply
Why it matters: Check the categories and suggested names, optionally run the dry run, click Confirm & Sort, and use Edit, then Undo last run, if you change your mind.
Key Features
Every item below comes from the README, FAQ or website; none has been tested by PromptQuorum.
- Categorization. A category and optional subcategory per file, in a More refined or More consistent mode, with named whitelists in which each main category can have its own allowed subcategories.
- Renaming. Suggested names for pictures and documents, and names built from embedded tags for audio and video, such as year_artist_album_title.mp3; a rename-only mode leaves files in place.
- Safety. A review table with file previews, a dry run showing From and To paths, a saved undo plan, and project-folder protection that skips Unity, Unreal, Godot, Blender, Git and common source-code projects.
- Memory. A local SQLite cache and a separate store of your approved decisions, both with clear and reset tools under Settings.
- Folders. Local folders, synced cloud folders, external drives and mounted network shares; on a NAS you first map a drive letter on Windows or mount the share on Linux and macOS.
- Languages. Interface languages include German, French, Spanish, Korean and Simplified Chinese; the category and filename language depends on the model, and Gemma 1.1 7B is English-only.
- Windows extension. An optional paid right-click add-on in the Microsoft Store starts cleanups from File Explorer.
AI Features and Where They Run
A language model reads file names, folder context and, where supported, file contents, and proposes the category and the new name; the model can be local or remote.
- Local text models. Gemma 3 4B IT, Gemma 1.1 7B and Mistral 7B are built-in choices, and you can register your own GGUF file; they run on an embedded llama.cpp runtime.
- Documents. PDF, Word, Excel, PowerPoint, OpenDocument and plain-text formats are read and passed to the selected model; legacy .doc, .xls and .ppt files are not supported.
- Pictures. Two built-in visual backends, Gemma 3 4B IT as the default and LLaVA 1.6 Mistral 7B, each needing a model file and a matching mmproj file; custom visual models work the same way.
- Audio and video. Only embedded metadata is used; the FAQ says whole videos are not sent to a vision model.
- Remote options. ChatGPT with your own OpenAI key, Gemini with your own key, or a custom OpenAI-compatible endpoint such as LM Studio, Ollama or a gateway; the README says keys are stored in the app's local config.ini.
- Learning. Approved decisions are reused as hints in later runs; the README says this does not train or change the model.
Data Flow and Privacy
The README says that with a local model your files, file names, images and metadata stay on your computer, no telemetry is sent, and an internet connection is needed only for a remote model. The categorization cache is a local SQLite database next to config.ini that stores directory paths, file names, categories and suggested names.
If you pick OpenAI, Gemini or a custom endpoint, the data goes to that service under your own account, so its policy applies. The only privacy text found for the app itself is a one-sentence statement linked from the Microsoft Store listing: it says no personal data is collected, stored or transmitted and that all processing stays on the device.
📌Note: These are the project's declarations, not audit results. The website has no privacy page, and the one-sentence statement does not mention the remote-model option, update notifications or model downloads. The README says a cache holds place names looked up from photo GPS data but not which service performs that lookup. PromptQuorum has not inspected network traffic.
Who Should Use It
- People with a cluttered Downloads, scans or screenshots folder who want suggested folders and names they can approve row by row.
- Privacy-minded users who prefer a local model and can spare the disk space for model files.
- Owners of external drives or NAS archives who will mount them first and sort in small batches.
What We Could Not Verify
- Quality and speed. Suggestion accuracy, run time and the effect of different hardware were not tested.
- Hardware needs. No RAM minimum for the local models and no model download sizes appear in the pages read.
- Undo. The README describes Undo last run as best-effort, skipping conflicts and files that changed; it was not tried.
- Privacy. Only a one-sentence statement exists; see Data Flow and Privacy.
- macOS. The builds are not Apple-signed, and the supported macOS versions differ between the README and the download page.
- Model licences. The licences of the downloadable models were not reviewed.
Competitors and Alternatives
The directory has two other local file and image tools with a review, neither of which renames and moves files the same way.
Hyperlink (Nexa AI)
- Platforms:
- Windows, macOS
- Price / license:
- Free / Proprietary
- Key difference:
- Searches and answers questions across your local files instead of renaming or moving them
Articles about Hyperlink (Nexa AI) (1)
- Hyperlink (Nexa AI) Review: On-Device File SearchUpdated September 12, 2026
PixlStash
- Platforms:
- Windows, macOS, Linux
- Price / license:
- Free / GPL-3.0
- Key difference:
- Image library manager with local AI tagging and search; covers pictures only
Articles about PixlStash (1)
- PixlStash Review: Self-Hosted Image Library Manager with Local AI TaggingUpdated October 17, 2026
Confirm each app's current price, license and platforms on its own page.
Frequently Asked Questions
What happens if I close AI File Sorter before confirming?
No files change, because analysis only prepares suggestions. The FAQ says saved results may be reused when you analyze the same folder again.
Can I use Ollama or LM Studio with it?
Yes, through the custom OpenAI-compatible option; the FAQ says Ollama is reached through its local server, usually at localhost:11434, and that LM Studio needs a vision-capable model for images.
Can it read scanned PDFs?
Not yet. The FAQ says scanned, image-only PDFs need OCR with another tool first, and that OCR support is planned.
Where are its settings and models stored?
Settings live in %APPDATA%\AIFileSorter on Windows, ~/Library/Application Support/AIFileSorter on macOS and ~/.config/AIFileSorter on Linux; downloaded models sit in a separate llms folder that can be moved.
Does it keep running in the background?
No. The FAQ describes it as a batch tool you open, run, review and close, and says there is no low-resource or background-throttle mode.
Verdict
AI File Sorter is a free, actively updated way to get reviewable folder and filename suggestions for messy files, with a local model as the default and an undo documented. Against that, the sources give no RAM figure for the local models, scanned PDFs and duplicate detection are not supported, the macOS builds are unsigned, the only privacy statement is one sentence long, and none of the quality claims could be checked here. It suits people who want to tidy a folder in small, reviewed batches; anyone who needs guaranteed offline behaviour should confirm it with their own network monitoring.
Sources
- AI File Sorter website — features, download, pricing, FAQ and changelog pages, checked 28 October 2026.
- AI File Sorter on GitHub — README, release notes and commit activity, checked 28 October 2026.
- AI File Sorter in the Microsoft Store — system requirements, update date and privacy statement, checked 28 October 2026.