Vev

The private browser where privacy is the default.

Vev is a desktop browser built from scratch: a Tauri 2 (Rust) shell embedding Chromium Embedded Framework (CEF) as the real multi-process rendering engine, with a Rust workspace for privacy, security, on-device AI, and downloading. Fingerprint resistance, WebRTC leak protection, AI phishing detection, ad/tracker blocking, always-separate private windows, and an encrypted vault are on by default — not paid add-ons.

• Website & docs: https://github.com/aryansrao/vev-website (deploys to Vercel)
• Community phishing feed: https://github.com/aryansrao/community-phishing-feed

---

Features

Privacy & security

• On-device AI (Huma). A real trained ML model (small MLP → ONNX, run by tract in pure Rust) scores every navigation, reads page content to catch brand-impersonation phishing, and adapts to your Allow/Block choices locally (a SEAL-style self-adapting layer). Plus on-device page summarization (Huma Read) and navigation prefetch (Huma Predict). No model provider, no telemetry.
• Fingerprint resistance. One fixed device profile (canvas, WebGL, audio, fonts, timezone, platform), injected into the main frame and web/service workers. The profile is per-OS so the JS values agree with Chromium's Sec-CH-UA-* headers (a cross-OS lie breaks real sites like Cloudflare Turnstile) — uniform across every install on the same OS, Tor-Browser style.
• WebRTC leak protection. Candidates limited to the public interface (no LAN IP in ICE); private windows remove WebRTC entirely.
• Ad & tracker blocking. Network-layer blocking (adblock-rust) + a live threat feed (URLhaus, OpenPhish, PhishTank, Spamhaus, Feodo) + a community-confirmed phishing feed with a real percentage per host. The Dynamic Island shows what was blocked and lets you allow per-site.
• Pop-up blocker. Script-spawned pop-unders (the kind sketchy streaming and torrent sites fire) are blocked outright, Brave-style — only a popup from a genuine click becomes a tab. Popups to threat-feed hosts are always blocked.
• Pre-open sandbox gate. A suspicious link is fetched first through a private proxy (Tor, else your custom proxy) into memory and analyzed — your tab never touches it until it's cleared. Your IP is never exposed by the probe.
• Multi-source deep scan. On-demand verification of a link against Google Safe Browsing, VirusTotal, URLhaus, ThreatFox, and AlienVault OTX, combined with the local model into one danger percentage — routed through the same private proxy so the scanners see a proxy exit IP, never yours. Opt-in; the everyday Guard stays fully on-device. Keys live only in the local config.
• Real private windows. Private browsing always opens a separate window with one ephemeral in-memory session and session-wide privacy switches.
• Encrypted vault. History, bookmarks, passwords sealed with AES-256-GCM; the key never leaves the machine.
• Network hardening. DoH (secure, no plaintext fallback), ECH, HTTPS-only, third-party cookies blocked.

Network routing

• Browser-wide proxy picker. Off / Tor (embedded Arti) / custom SOCKS5 or HTTP proxy (e.g. a Mullvad endpoint). Chosen on the private-window home page. Because CEF's Alloy runtime refuses per-context and runtime proxy changes, a proxy is a launch-time switch — the picker persists your choice and offers one-click Apply & relaunch. --vev-tor-all still works from a terminal.

Downloads

• Media downloader (yt-dlp + ffmpeg). From any supported site (YouTube, and 1800+ more): pick format and quality, extract MP3, or Play online (streams instead of saving). yt-dlp/ffmpeg are used from your PATH if present, else fetched once into app-data — nothing bundled into the installer.
• High-speed segmented downloads (parallel range requests) with an IDM-style manager: live progress, pause/resume/cancel, show-in-folder, and a clear-from-list control.
• Torrent client (librqbit). .torrent files and magnet: links are auto-detected — including by content-sniffing the downloaded bytes, so a torrent served with a movie-title filename still routes to the engine instead of saving as a file. Pause/resume/remove, live progress, streaming download.

Everything else

• Content-script extensions, the command palette, find-in-page, zoom, tab drag-reorder, ⌘L, search suggestions, and every keyboard shortcut you expect.

Build & run

Requires Rust (stable), the CEF binary distribution in ~/.local/share/cef, and the bundle-cef-app helper (from a cef-rs checkout). Then:

make run      # build + bundle + launch
make test     # cargo test across the workspace
make bundle   # assemble the app bundle without launching
make autotest # in-app runtime self-test suite
make train    # retrain the Huma phishing model (ONNX)
make help     # list all targets

CEF needs the assembled .app bundle to spawn its helper processes, so a bare cargo run cannot launch the browser — always use make run/make bundle. First launch takes ~15 s to bring up CEF and the first tab.

Platforms

macOS (Apple Silicon) is the verified, day-to-day build. Windows and Linux are supported by a cross-platform window-embedding layer (src-tauri/src/platform.rs): the CEF child view attaches via the native handle for each OS (NSView on macOS, HWND on Windows, X11 window on Linux, resolved through raw-window-handle), and view show/hide/detach use the matching platform API. All three OSes compile in CI (CEF auto-downloaded), so the platform code is build-verified; Windows and Linux still need on-hardware runtime testing (Linux is X11-only — Wayland can't host a CEF native child).

Do not re-sign the macOS bundle with codesign --deep — it corrupts CEF's nested signatures; rebuild via make bundle, which signs correctly.

Architecture

• src/ — the shell UI (chrome, omnibox, Dynamic Island, internal pages, download popup, online player), plain HTML/CSS/JS in the Tauri webview.
• src-tauri/ — the Rust app: CEF lifecycle, tab manager, commands, downloads, media, sandbox, and the render-process helper (vev_helper) that injects fingerprint resistance into every V8 context.
• crates/ — the workspace: huma (AI: guard/content/adapt/model/read/ predict/intel), vev-fingerprint, vev-network, vev-blocklist (incl. the community feed), vev-storage, vev-tor, vev-download, vev-media (yt-dlp/ffmpeg), vev-torrent.
• scripts/train_guard/ — the Huma model training pipeline (features come from the same Rust extractor used at inference, so train/serve are identical).

Privacy model

Nothing about your browsing leaves your device by default. History, bookmarks, and passwords are encrypted locally; the AI runs in-process with no model provider; the only optional outbound privacy request — a community phishing report — is off by default, manual, and sends just a hostname + the AI score (never full URLs or history). Downloads save plainly to ~/Downloads (and torrents to ~/Downloads/vev-torrents); they are not encrypted, and Vev does not claim otherwise.

Honest limitations

• A few fingerprint surfaces (CSS-media screen size, local font enumeration, residual headless signals) still need engine-level work — documented, not hidden.
• Proxy routing (Tor/custom) is browser-wide and applied at launch: the embedded engine refuses per-context and runtime proxy changes, so switching needs a relaunch. Per-tab/per-window proxying isn't possible.
• Extensions are the content-script subset; background pages / popups / chrome.* APIs aren't supported by the embedded engine (Vev says so up front).
• The media downloader needs yt-dlp for streaming sites and ffmpeg for MP3/format conversion (used from PATH or fetched once). YouTube can throttle yt-dlp to few formats without a po-token plugin.
• YouTube in-stream ads are best-effort (they're muxed with the video).