GUIDE · PRIVATE SEARCH

Ask your Mac, not someone else's server.

Shlexball's AI runs entirely on your Mac, finding lost files like your GitHub and Stripe recovery codes in seconds — and your files never leave your Mac.

26 September 2026 · 30 second video

The searches are real, recorded on an Apple silicon Mac with the model running on the Mac. The recovery codes are made-up samples.

Watch it on YouTube · More how-to videos

How accurate is it?

"Executed tests passed" measures whether answers actually work. Each test is a plain-English search request, written by our test generator: the model's answer is run by Shlexball's search engine on a prepared folder holding the right files plus look-alike decoys. A test passes only if the search returns exactly the right files — nothing missing, nothing extra — or correctly refuses a request it can't do safely. Nothing is graded on "sounds right." Shlexball's 1.5B model passes 308 of 312 requests it never saw in training (98.7%), scored with plurals: the singular and plural step the app applies before every search. Updated 4 October 2026.

See the full FindBench results

FAQ

Does Shlexball upload my files?

No. Shlexball's AI and search engine run entirely on your Mac. Document contents, file lists, and previews are never uploaded. After the model download (about 3.1 GB), search works offline within your free searches or paid period. Sharing search diagnostics is off unless you choose to opt in. See the privacy policy.

Why not just ask a cloud AI to find my files?

A cloud AI runs on someone else's server, so what you ask it — and, with many tools, file names or document contents — is sent there, where you don't control who reads it or how long it's kept. Shlexball's AI runs on your Mac: your files never leave it, and your searches stay on it unless you choose to share diagnostics.

How can a 1.5B model be accurate?

By doing one job well. Shlexball's model is specialised for file search and graded with executed tests: its answers are run by the real search engine on folders with look-alike decoys, and only exact results pass. It passes 436 of 479 tests (91%), including 90% of requests it never saw in training.

▶ DOWNLOAD

Free to try with 30 searches. macOS 15 or later. Apple silicon. 16 GB memory.