What goes in.
What goes out.
AnoVeil talks to our services to check who you are and what you have paid for. It never sends them your documents. This page is the full list of what travels on each path.
The engine is local
Document processing runs in a local engine bound to 127.0.0.1. The interface is embedded in the application and loads nothing from a CDN.
The models are bundled
Detection models ship inside the application and offline mode is forced whether or not the cache is present. A missing asset falls back locally instead of fetching one.
A test enforces it
Every non-loopback socket is blocked, then a scan, an anonymise and a restore are driven through. If any reaches out, the build fails. A second test proves the block itself bites.
Four things.
Server-validated
Account, subscription and device access are checked by AnoVeil’s services, not trusted from whatever the local application claims.
Device authorization
A device must be authorized before protected functionality is enabled — per device, not per install.
Revocable
Authorized access can be revoked from your account portal. Once revoked, it is not treated as valid again.
Fails closed
If a check fails, an identity mismatch occurs, or access has been revoked, the application stops treating the device as authorized rather than carrying on.
As little as
the service needs.
Where the rest
of it lives.
%LOCALAPPDATA%\AnoVeil Your codebook, your vocabulary, and an unfinished review if you closed mid‑way. On first run you choose whether the codebook is password‑protected or stored as‑is. Uninstalling leaves all of it in place.
AnoVeil is Windows‑only today. macOS and Linux are not supported yet.
What it misses.
Detection is strongest on structured values and weakest on people’s names, the category that matters most. We publish the per‑category figure rather than a blended average, because adding easy categories lifts an average without the hard case moving.
The figures are recall — the share of the person names present that AnoVeil found — measured over a 60‑document corpus, most recent run. Precision on names is higher, at 0.95: when it marks a name, it is almost always right. Across runs recall has ranged from 0.76 to 0.85, and we publish the latest rather than the best.
Review every document before you send it. AnoVeil shows what it found and lets you add what it did not — the review step exists because detection is a safety net, not a guarantee.
If you believe you have found a security vulnerability in AnoVeil, please report it privately. Include enough detail for us to understand and reproduce the issue, and please give us a reasonable opportunity to investigate and address it before disclosing publicly.