OfflineMask

Mask and restore sensitive data, entirely on your device.

▶ Watch the introduction video

OfflineMask is a desktop app for preparing sensitive text before pasting it into an external AI tool. You paste text, the app replaces names, account numbers, phone numbers and other private values with consistent aliases on your machine, you review the result, copy it to the AI of your choice, and paste the reply back to restore the original values. Nothing leaves the device: there is no server, no telemetry and no automatic update check. The first language focus is Vietnamese banking workflows; the detection pipeline is designed so that other languages and models can be added as packs.

Status: 2.0.0 is the first public release. Pasted text, spreadsheets (XLSX, XLS, ODS, CSV), documents (DOCX, ODT) and plain-text files are masked, reviewed and restored; a PDF that contains text is masked to a Markdown copy. Once a passcode is set, history is kept in an encrypted vault on the device. On macOS and Windows, Ẩn nhanh masks or restores text selected in any application. Scanned PDFs and images are not implemented. Only macOS on Apple Silicon has been built and tested by hand; Windows x64 builds in CI and awaits manual qualification; Linux is untested. Builds are not yet signed or notarized. Detection is rule-based plus an optional model pack and has documented misses; always review the result before sharing it. See the changelog for what each release contains, the product requirements for the intended scope and product direction for the decisions behind the current experience.

How it works

• Built-in rules recognise emails, phone numbers and labelled banking identifiers (accounts, national IDs, passports, tax codes, transaction and customer codes) without any model files. The app works offline from the first launch.
• An optional recognition pack adds contextual categories: names, addresses, organizations, birth dates, card and login details. The pack is downloaded on demand from Settings (712 MB transferred, 1.14 GB installed), verified by size and SHA-256, and runs locally through ONNX Runtime inside a web worker. Startup and processing never touch the network.
• Review and aliases. Every finding can be kept or hidden; a manual selection can be hidden too. Aliases are consistent within a session, so the AI's reply can be restored exactly. Without a passcode, private sessions live only in memory and disappear when the app closes; with one, they are sealed in a local vault (Argon2id-derived key, XChaCha20-Poly1305, Decision 0062). A forgotten passcode cannot be recovered.
• Files and other apps. Spreadsheets and documents are masked with their formatting kept and a review of every hidden value before saving; the AI's returned file is restored the same way. Ẩn nhanh (Decision 0096) puts a small button beside text selected in any application and replaces it in place; it reads only the selection, on the user's click.

Detection quality is evidence-based and recorded, not promised: the shipping pack reproduces all 398 recorded synthetic outcomes of its evaluation, which still include known misses and over-masking. Read the integration evidence before relying on it.

Getting started

Builds for each release are attached to the GitHub releases by the release workflow: macOS Apple Silicon and Windows x64 with native recognition, macOS Intel and Linux x64 with WASM recognition. They are not yet signed or notarized (release notes). To build from source, install mise and the Tauri 2 prerequisites for your OS (on macOS the Xcode Command Line Tools suffice), then:

mise trust
mise install
mise exec -- pnpm install --frozen-lockfile
mise exec -- pnpm -C desktop desktop:dev

mise.toml pins Node, pnpm and Rust; mise exec -- applies them even when mise is not activated in your shell. The repository is a pnpm workspace with two packages: desktop/ (the Tauri app) and landing/ (the product website); pnpm -C desktop … and pnpm -C landing … run a package's scripts from the root, and pnpm lint at the root covers both. Downloading build dependencies needs a network connection; the application itself does not.

Command	Purpose
mise exec -- pnpm -C desktop dev	Browser preview at localhost:1420 without native APIs (no pack download, built-in rules only).
mise exec -- pnpm -C desktop desktop:dev	The desktop app with hot reload. Recognition runs in WASM (slow on long input).
mise exec -- pnpm -C desktop desktop:dev:native	The same with the native recognition helper (macOS arm64, Windows x64): release-speed masking while developing.
mise exec -- pnpm lint	ESLint and Prettier; pnpm format rewrites.
mise exec -- pnpm -C desktop typecheck	Vue, TypeScript and Vite configuration.
mise exec -- pnpm -C desktop test	Fast tests: detection rules, transformations, sessions, storage boundaries, the pack contract. No model files needed.
mise exec -- pnpm -C desktop build	Production frontend without model weights.
mise exec -- pnpm -C desktop check:native	rustfmt and Clippy for the Rust layer; pnpm -C desktop test:native runs its tests.
mise exec -- pnpm -C desktop test:models <pack root>	Run the real recognition pack against the committed synthetic fixtures (see packs/README.md).
mise exec -- pnpm -C desktop desktop:build	Desktop bundle for the current platform. macOS arm64 uses the verified native-inference path.
mise exec -- pnpm -C desktop desktop:build:native	Native-inference bundle on macOS arm64 or Windows x64 (Windows awaits qualification, Decision 0077).

Only macOS on Apple Silicon has been built and tested so far. The CI workflow compiles the native layer on Linux and Windows as an advisory check, and a separate workflow builds and smoke-tests the Windows native package; Windows becomes qualified after the manual checklist in Decision 0077.

Replacing the models

Everything the app knows about its models is one file under desktop/packs/. The TypeScript loaders, the Rust installer and the evaluation helpers all read it, so a different model, or a pack for another language, is a change to that file plus the files it pins. packs/README.md walks through the contract (inputs, outputs, labels, offsets) and the checks a replacement has to pass: pnpm test for the code around the models and pnpm test:models for the models themselves.

Product landing page

The standalone website lives in landing/. Run mise exec -- pnpm -C landing dev for a local preview. It has its own Cloudflare Workers build and manual deployment workflow, with reusable content for the future Astro/EmDash product page.

Repository layout

Path	Contents
desktop/src/core/	Portable logic: detection rules, model policy, pack loading, aliases, sessions, restoration. No Vue or Tauri imports.
desktop/src/core/detection/	The detection pipeline. pack.ts reads the pack, recognition.ts loads it, policy.ts merges model and rule evidence.
desktop/src/adapters/	The only modules that talk to Tauri: the detection worker bridge, pack status and downloads, the website link.
desktop/src/workers/	The detection web worker.
desktop/src/composables/, desktop/src/pages/, desktop/src/components/	Vue state and screens.
desktop/src-tauri/	Rust: pack download and verification, exact tokenizer offsets, opt-in native inference.
desktop/packs/	Recognition pack definitions.
desktop/tests/	Node test runner suites and synthetic fixtures.
desktop/scripts/	Build tooling at the top level; frozen evaluation experiments in subdirectories (index).
docs/decisions/	Architecture decision records, numbered.
docs/evaluations/	Evaluation protocols and results; large raw outputs are kept out of git.
brand/	The mark (a brush swipe with two eye holes), its generator, sketches and icon sources, and theme.css, the colour and type tokens shared by the app and the landing page.
landing/	The product website: a static page for Cloudflare Workers, with its own README.

docs/development.md has the detailed setup notes, module boundaries and the history of evaluation tooling.

Privacy principles

Only synthetic data belongs in this repository, in issues and in screenshots. The app stores no private text, collects no corrections and sends nothing anywhere; the sample history on disk holds references to public fixtures and mask/keep choices only. Changes that would relax any of this need a recorded decision. Vulnerabilities are reported privately as described in SECURITY.md.

Contributing and license

See CONTRIBUTING.md for conventions and the checks to run, and CODE_OF_CONDUCT.md for expectations. Agents working in this repository follow AGENTS.md.

OfflineMask is source available under the O’Saasy License Agreement. Copyright © 2026, Phong Phan. Third-party code and the recognition models keep their own licenses, recorded under docs/licenses/ and in the pack attribution. See the license decision.