Dev preview Built local-first for a hackathon

PrivacyMyst reads each page the way a data broker would, warns you the instant a site starts building a profile, and blocks known trackers, malware and phishing. The analysis runs entirely on your machine — the text you read never touches a server.

Local AI — no cloud inference Signed blocklist, verifiable MIT open source No accounts, no email
Scroll to watch it work
Live · simulated

Watch a profiling attempt fall apart.

Open a page that wants your data and PrivacyMyst scores the risk on-device, then severs every tracker, fingerprinter and malware host before it phones home. Scroll to run it.

Scan idle
Page secured · you’re invisible
health-insurance-quotes.example
PrivacyMyst 0
0 Risk score

Scanning…

On-device classifier · WebGPU

Blocked in real time0

Illustrative recreation of the extension UI. Page text is classified locally and discarded — nothing is uploaded.

How it works

Three layers, all running inside your browser.

01

It reads the page like a profiler would

A local model scores every page for sensitive context — health, finance, identity — and tells you the moment a visit could become a profiling event. The text never leaves the tab.

02

It blocks the network behind the page

A curated, license-clean list of trackers, malware and phishing domains is enforced everywhere. On a page the AI flags as sensitive, blocking escalates to a stricter tier scoped to that site only.

03

It tells you exactly why

Every block is labelled with the reason and the source list, so you can see precisely which broker or malware host was about to receive your data.

What's inside

A full privacy toolkit — not a single trick.

Six defenses that layer together. The first two are live and verifiable today; the rest are in active development.

AI Deep Dive

A local language model and classifier flag profiling and manipulation risk on the page you're reading.

On device

Risk-adaptive blocking

Loose baseline everywhere, scorched earth on sensitive pages — driven by the page risk score, not a fixed rule.

Signed feed

Link Guard

Checks links on hover with zero network calls and adds a short cool-down before risky clicks.

Zero network

MailGuard

Reads rendered webmail locally to flag sender spoofing, risky attachments and known campaign patterns.

On device

Fingerprint defense

Bionic Blur softens canvas, audio and motion signals so trackers can't stitch your sessions together.

Per origin

Honeypot & noise

Poisons tracker parameters and injects believable noise, so the profile that gets built is simply wrong.

Active
Why you can trust it

Privacy software you can't verify is just a promise.

This one is built to be checked — by you, line by line.

The AI never phones home

Models run in your browser over WebGPU and WebAssembly. There's no server to send your page text to — because there is no server.

Updates are signed, not trusted

The blocklist is delivered as data, never code, and every update carries an Ed25519 signature the extension verifies before applying. A broken or hostile server can't poison your install.

It fails safe

If an update is missing, stale or unsigned, the extension keeps the last good list. The worst a bad update can do is block too much — never unblock malware or leak data.

every update is verified on your device
fetchbundle.signed.json
verifyEd25519 sigreject if forged
checkversion > lastreject rollback
checksize sanereject anomalies
applyblock rules onlycan't redirect

Threat feeds are open-source and license-clean. Overlapping lists are deduplicated so the same bad domain is never counted twice.

Open source · MIT

Don't trust us. Read the code.

Every layer — the on-device models, the signature checks, the blocklist pipeline — is public and auditable. Fork it, build it, break it, send a PR. Privacy you can't inspect isn't privacy.

0%
Of inference runs on your device
0
Bytes of page text uploaded
0k
Blocked domains in the signed list
Ed25519
Signature on every update
Get the extension

A developer preview for Chromium browsers.

No sign-up, no account, no email. The package ships its full on-device AI models, so it builds from source rather than a casual download.

01

Build from source

Clone the repo and run the build script: npm run build.

02

Enable developer mode

Open your browser's extensions page and switch on developer mode.

03

Load unpacked

Choose Load unpacked and select the build output folder.

PrivacyMyst

PrivacyMyst — Developer Preview

Local page analysis & the signed blocklist are live and verifiable today.

Clone the repo