What A.Redact collects (and what stays in your browser)

A.Redact runs in your browser. Your PDFs stay on your device. The only things that leave the browser are (a) extracted page text sent to an AI provider when you turn the optional AI detection on, and (b) anonymous usage events for the Research Preview.

• The PDF you upload (parsed locally by pdfjs-dist).
• Every detected entity (names, SINs, phone numbers, addresses, etc.) and their character offsets.
• The redacted output PDF (rewritten locally by pdf-lib and downloaded directly).
• Any custom dictionary entries or AI prompts you add in the extension popup (stored in chrome.storage.local on your machine; never synced to a server).

This is non-negotiable in the codebase: the source PDF binary is never uploaded by A.Redact to anyone, regardless of which other features you enable.

On both the web app and the browser extension, you can opt in to an "AI detection" pass that catches names, free-form addresses, organisations, and anything described in your custom AI prompt that the regex rules miss. When you enable it:

• The extracted text of each page (one page per request) is POSTed to /api/llm/extract on our server, then forwarded to the configured LLM provider (default: Cohere; fallback: Anthropic).
• The PDF binary is not sent. Detected values from earlier passes are not sent. File names, file sizes, and your IP-derived identity are not attached.
• Our route does not log or transmit the request body, the response body, or any content derived from either. A file-header banner in the route source enforces this.
• We do record an anonymous operational event for each request (which provider answered, how long it took, whether it succeeded or failed) so we can monitor the endpoint's health. The event carries no document content and no entity values; the event identifier is a short, irreversible bucket derived from your IP address (so repeat requests from the same caller cluster together) and your raw IP is not stored.
• The provider's own retention policy applies to the text while it's with them. Cohere and Anthropic do not retain API content for training by default.
• Per-IP rate limiting is in place to prevent abuse of the public endpoint.

For the duration of the Research Preview we collect more behavioural data than we will once A.Redact reaches general availability. This is an explicit trade we made so we can learn what works and what does not before scaling. The full scope on the website:

• PostHog Cloud (EU, Frankfurt): pageviews, autocaptured clicks and form interactions, session replays with every input field masked (maskAllInputs: true), persistent anonymous distinct IDs, person profiles, heatmaps, dead-click tracking, performance metrics, unhandled-exception capture, browser console output our app emits during your session, and IP for country / region inference.
• Browser extension: same PostHog project tagged $lib: aredact-extension, using a per-install distinct ID stored in chrome.storage.local. Events are coarse: file size buckets, redaction counts (no values), AI toggle changes, popup interactions, waitlist signups. No file names, no entity values, no document content in any event. When you click a link in the extension popup that opens our website, the extension's anonymous distinct ID is passed to the website so your activity on the two surfaces can be linked into a single anonymous profile.
• Sentry (US): unhandled exceptions and stack traces, scrubbed via src/lib/pii-scrub.ts for SIN / credit card / phone / email / postal-code / filename shapes before transmission. Session replays disabled. Bug reports and product-feedback messages submitted via the /feedback form (and the in-popup "Send feedback" form in the extension) are also stored in Sentry as User Feedback entries, along with the email address you provide if you choose to leave one. The same submission is also transmitted to Resend (US) as a one-shot transactional email so the project owner is alerted in real time. The email contains your message body, your optional name, and your optional reply-to email; nothing else.
• Vercel Speed Insights (US): Core Web Vitals (LCP, INP, CLS, TTFB) only. No identifiers.

Compliance trade-offs are logged in our public COMPLIANCE_RISKS.md register (entries R-009 and R-012). We will tighten this scope and ship a Quebec Law 25-compliant consent banner before general availability.

If you join the waitlist (web form or in-extension popup), we send your email address, your marketing-email opt-in flag, and a tag identifying which surface the signup came from (website or extension) to MailerLite (EU, Vilnius). We use it only to email you when production beta invitations open up. You can unsubscribe at any time using the link in any email we send.

When you sign in, we store your email address and (if you choose) your display name and preferred interface language in our database. We do not keep any record of which PDFs you redacted, what content they held, or what you produced. Sign out at any time from the Account page; you can also delete your account permanently from there, which removes the profile row and your analytics history.

On the website, PostHog sets cookies and writes to localStorage for its persistent distinct ID. Vercel Speed Insights writes a short-lived cookie. The browser extension uses chrome.storage.local for your settings (AI toggle, dictionary, AI prompt, blocked-sites list, the PostHog distinct ID). None of that is synced to a server by us.

A.Redact is not directed at children under 13 and we do not knowingly collect personal data from them.

We will bump the effective date at the top of this page when we change scope. Material changes (new third-party service, new telemetry category, account creation going live) get an additional announcement on the product itself.

Questions, requests for deletion, or anything else: info@anaris.xyz. Anaris Labs is the legal entity behind A.Redact.