Searching Mac files with plain language
Shlexball arrives as a Mac application that lets you find files by describing them in plain English. Instead of remembering exact filenames or digging through folder hierarchies you type what you recall—a phrase a date a document type—and the app returns matches. The search runs entirely on your machine. No files leave your Mac. No cloud uploads occur. The project targets users who need to recover a contract a presentation or a recovery code without leaving the desktop or switching to a browser window.
Local LLM powers on-device searches
The core of Shlexball is a local large language model that processes your description and turns it into a file search. During initial setup about 3.1 GB of model files download over the internet. After that cached searches run completely offline on Apple silicon. The model reads file names and inspects the text inside PDFs Word documents and other supported formats. It never modifies moves or deletes anything it only reads. The app requires macOS 15 or later and 16 GB of memory for the model to run smoothly. A 154.3 MB client download is available with a SHA-256 checksum of 34e065c2…fe4d81 for verification.
Benchmarks put Shlexball ahead of cloud models
The team behind Shlexball created FindBench a custom evaluation set of 312 plain-English file-search requests the model had never seen during training. Each request gets turned into a search against a test folder containing the right files plus decoys that look similar. A test passes only when exactly the correct files return nothing missing nothing extra. On those 312 unseen requests Shlexball passed 308 tests. Gemini 3.8 Flash running on Google's servers passed 295. The lead is 13 tests. The benchmark also checks response time. Shlexball averages 0.4 seconds per answer on an M4 Max MacBook Pro fully on-device. Gemini 3.5 Flash-Lite through Google's API averages 0.6 seconds when network latency is included.
How the search works in practice
When you type a description the app expands each topic word into its singular and plural forms before sending the query to the model. This means a search for "invoice" also catches files containing "invoices." The model then searches both filenames and the internal text of documents. Spotlight on the other hand matches names and indexed words only so a file you can describe but cannot name often slips past it. Shlexball's approach is especially useful for finding documents by content such as "the lease PDF from last spring" or locating recovery codes for GitHub or Stripe. The example in the launch film shows a user typing that exact phrase and the app returning three matching files.
System requirements and pricing
Shlexball is available as a Mac app with 30 free searches no account required. After the free tier pricing is 9 USD per month or 59 USD per year. The app sends you a link you can email to yourself for later installation. No ongoing email list is kept. The software runs on macOS 15 or later on Apple silicon machines with at least 16 GB of memory. First-time users download the 3.1 GB model during setup after which all searches operate from cache on the device. The app only reads files it never changes moves or deletes them. A short terminal installation option is also documented for power users who prefer the command line.