Sci

Appearance

The Privacy Guarantee ​

Sci's core promise: your real name, email, and identifying information never appear in a prompt sent to an AI provider.

This is verifiable. You don't have to trust us.

Run the demo ​

bash
node demo/privacy-demo.mjs

This takes under 10 seconds and shows you exactly what happens to a message containing your real identity before it would reach an AI provider.

Expected output:

Step 1: What you typed
Hi, my name is Casey Zandbergen...

Step 2: What the AI provider would receive
Hi, my name is [PERSON_1] and I'm based in [PLACE_1]...

Step 3: Token map
  [EMAIL_1]  ← "casey.zandbergen@gmail.com"
  [PERSON_1] ← "Casey Zandbergen"
  [PLACE_1]  ← "Tulsa, Oklahoma"
  ...

Step 4: Verification
  ✓ Real name absent from outbound text
  ✓ Email absent from outbound text
  ...

✓ All checks passed.

How it works ​

The anonymization pipeline runs in four layers, in order:

Layer 1 — Regex (works for everyone, day one)

Emails, URLs, phone numbers, social handles. Pattern-matched, no training required. A brand-new user gets 100% coverage immediately.

Layer 2 — compromise.js NER

Named entity recognition for people, places, and organizations using a pre-trained English model. "John Smith from Seattle at Acme Corp" gets caught with no history.

Layer 3 — Custom entities from your identity_facts

Project names, company names, and other personal entities learned from your history. The longer you use Sci, the better this layer gets. On day one, seed it with sci import --claude.

Layer 4 — CamelCase detection

Structural detection of compound proper nouns (OpenClaw, CrossTimbersFarm, ElevenLabs) regardless of history.

The token map ​

Every detected entity gets a token: [PERSON_1], [EMAIL_2], [URL_3]. The mapping is stored in process memory only, for the duration of the MCP server process.

It is never:

  • Written to disk
  • Written to the database
  • Sent to any external service
  • Persisted across server restarts

Use session_inspect to audit what has been masked before any outbound call.

Session feedback loop ​

If a new project name (like Threadline) gets caught in call 1 of a session, it's automatically in the session's entity set and proactively checked in all subsequent calls — even in contexts where NER might not catch it.

Progressive promotion ​

If an entity appears in 3 or more separate calls, it gets automatically promoted to identity_facts and will be caught in every future session via Layer 3.

What Sci does not guarantee ​

  • Perfect NER coverage — named entity recognition is probabilistic. Uncommon names or highly context-dependent entities may slip through, especially for new users with no imported history.
  • Protection against inference — if you tell the AI "I'm working on an iOS wardrobe app" without naming it, the AI can infer context even without the project name.
  • Post-response handling — Sci deanonymizes the AI's response before showing it to you. What the AI does with its own context window is outside Sci's scope.

The anonymization engine is conservative: it's better to mask something generic (like a city name) than to miss something sensitive. Use session_inspect if you need to audit before sending.