@latten/redact

Keep your secrets and your users' data on your machine — without slowing down.

You went all-in on AI. @latten/redact is the throttle you keep your hand on: it detects secrets and PII on-device, replaces them with reversible tokens before a single value reaches the model, and keeps a values-free tally of what it held back. Everything runs in-process. Nothing leaves your machine. You keep moving at full speed and still sleep at night.

npm install @latten/redact

Free: protect your own coding-agent loop

Detect and redact secrets + PII in any text, locally. The model still gets a coherent prompt (same entity → same stable token), and you restore the exact values on the way back — do no harm.

import { redact, rehydrate, flag, newFlagState } from '@latten/redact'

const { text, map, findings } = redact(
  'email [email protected] about key AKIAIOSFODNN7EXAMPLE',
)
// text:    "email [EMAIL_1_9f3a1c20] about key [SECRET_1_9f3a1c20]"
// findings: [{ type: 'email', count: 1 }, { type: 'aws_access_key', count: 1 }]  ← types + counts, never values

const answer = await callModel(text) // the value never left this process
const restored = rehydrate(answer, map) // exact round trip

// Calm, in-session: explain once per kind, then quiet down. Never alarmist.
flag(findings, newFlagState()).lines.forEach((l) => console.log(l.text))

• redact / rehydrate / reveal — reversible, do-no-harm tokenization. reveal re-adds a single value if the result truly needs it (we recommend against it).
• detectSecrets / detectPii — values-free { type, count } summaries.
• createLedger / record / summarize — a local audit tally: type, count, and time only. No values, no value-hashes.
• flag — calm in-session illumination you can show through your agent.

Paid: protect your production app

protectApp() points the same engine at your app's live LLM calls, gated by a signed entitlement verified offline — so a Latten outage can never block a customer request. Redaction runs in-process; nothing here makes a network call on the request path.

import { protectApp } from '@latten/redact'

// LATTEN_ENTITLEMENT comes from your Latten for Developers account.
const guard = protectApp({ onFlag: (lines) => log(lines) })

const { text, map } = guard.redact(userPrompt) // fresh map per request — tokens never bleed across users
const answer = await callModel(text)
return guard.rehydrate(answer, map)

Production-app protection is part of Latten for Developers (Solo and Team). Free covers your own coding-agent loop forever.

Privacy model

Values-free by construction. Only { type, count } summaries are emit-safe (WireSafe); anything that can hold a raw value (the rehydration map) is LocalOnly and never crosses the wire. Detection is honest, not magic — the structural guarantee is that values stay on your machine, and we fail closed (over-redact when unsure) rather than leak.

Looking to keep your AI spend in view with the same on-device, values-free posture? That's our sibling package, @latten/cost — local detectors for redundant calls, stuck loops, and oversized or uncached context.

You moved fast with AI. Latten helps you stay in control of it. See what your AI costs — and where your data goes: Latten turns the AI calls across your company into one live graph — cost, reach, and PII exposure, attributed and observed from real traffic. One prompt to your coding agent, a PR you review, live in minutes. latten.io