prompt-scrub

Built by the Nano Collective — a community collective building AI tooling not for profit, but for the community.

prompt-scrub is a local-first utility designed to strip identifying content out of prompts and messages before they hit any cloud LLM.

It maps sensitive data (emails, secrets, paths, URLs, phone numbers) to stable placeholders locally, allowing you to rehydrate the model's responses back to their original forms securely.

What it is / What it is not

prompt-scrub reduces identity leakage at the content layer. It is partial defence, not anonymity.

What it does:

• Detects and replaces common identifying content (emails, paths, phone numbers, secrets, URLs) before your prompt leaves your machine.
• Maps each value to a stable placeholder so the model's response can be rehydrated locally.
• Gives you inspect (table of detected entities with confidence and method), diff (visual before/after), and --min-confidence <0-1> on scrub/inspect to drop anything below a threshold — so you can see what was detected, what was missed, and exactly what will be sent before you commit to sending.

What it does not do:

• It does not make you anonymous. A semantically identifying question (a niche bug only you have, your private codebase, your financial situation) remains identifying after scrubbing.
• It does not address stylistic fingerprinting — the way you phrase things goes out unchanged.
• It does not operate at the network or key layer. Your IP address, request timing, and headers are outside its scope.
• Detectors can and do miss things. Always review the output before sending.

[!IMPORTANT] A user who believes this tool makes them anonymous is worse off than one who never used it — they stop reading their prompts and trust the defaults. Always use inspect first to see what the tool actually found.

Read the full Threat Model for a complete breakdown of what is and is not defended.

Quick Start

Install globally to use the CLI:

npm install -g @nanocollective/prompt-scrub

Or install as a dependency in your Node.js project:

npm install @nanocollective/prompt-scrub

Recommended: Inspect first

Before scrubbing, run inspect on a real prompt to review what the tool detected before sending your prompt:

echo "My email is [email protected] and I work at /Users/alice/projects. My phone is +44-7700-900999." \
  | prompt-scrub inspect

Detected entities:
  [Email]    [email protected]                   → «Email_1»  (chars 12-26, confidence 0.95 exact-pattern)
  [Path]     /Users/alice/projects.           → «Path_1»   (chars 41-63, confidence 0.80 structural)
  [Phone]    +44-7700-900999                  → «Phone_1»  (chars 76-91, confidence 0.90 structural)

No session written.
Hash: 7e5eea933db987e10e10e259ebcfea9d3250d8a68925fd9360f515e3a4bfbba9

Each entity carries the confidence the detector assigned it: an exact vendor key pattern scores 0.99, a capitalised-word name guess only 0.5. Pass --min-confidence <0-1> to scrub or inspect to drop everything below a threshold. See Confidence & Tiered Detection.

The hash is deterministic — the same prompt always produces the same hash, so you can verify cache stability across runs. Once you are satisfied with what inspect shows, run diff to see the replacements in place, then proceed with scrub.

echo "My email is [email protected]" | prompt-scrub diff --no-color

- My email is [email protected]
+ My email is «Email_1»

Usage Examples

CLI: Scrubbing text

echo "My email is [email protected]" | prompt-scrub scrub
# stdout: My email is «Email_1»
# stderr: Session ID: 6f1c2b90-...
# stderr: Scrubbed: 1 entity (1 Email)   — pass -q/--quiet to suppress the summary

# Only scrub what the detectors are confident about
echo "My email is [email protected]" | prompt-scrub scrub --min-confidence 0.9

CLI: Encrypting session files at rest

Local session files map placeholders like «Secret_1» back to your real secrets. If your laptop is stolen or another process on your account reads ~/.config/prompt-scrub/sessions/, those secrets leak. Enable encryption to keep them protected on disk:

// ~/.config/prompt-scrub/config.json
{ "encryptionEnabled": true }

Then pass a passphrase via the PROMPT_SCRUB_KEY environment variable when running scrub/rehydrate/sessions:

# Read your passphrase from a secret manager; do NOT echo it inline — it will
# land in your shell history.
export PROMPT_SCRUB_KEY
read -rs PROMPT_SCRUB_KEY   # enter interactively
echo "sk-1234567890abcdefghij" | prompt-scrub scrub

If PROMPT_SCRUB_KEY is unset and stdin is a TTY, prompt-scrub will prompt for the passphrase interactively (with input muted). Rehydrate on an encrypted session also requires the same key.

Note: prompt-scrub only reads the file rehydrate was asked about. If you only ever use scrub, the output (the scrubbed prompt) is safe to send anywhere regardless of encryption; encryption protects the session file on disk.

To migrate existing plaintext sessions after enabling encryption:

PROMPT_SCRUB_KEY=… prompt-scrub sessions encrypt

CLI: Watch clipboard

# Monitor clipboard and automatically scrub sensitive data
prompt-scrub watch --clipboard

# Watch a file
prompt-scrub watch --file prompt.txt --once

# Preview changes without writing anything
prompt-scrub watch --file prompt.txt --dry-run --once

See the CLI Reference for all watch options and platform requirements.

CLI: Run a transparent scrubbing proxy

# Point your SDK at the proxy instead of the upstream API.
prompt-scrub proxy --target https://api.openai.com --port 8080 &
export OPENAI_BASE_URL=http://localhost:8080/v1

# Any request your app sends to OpenAI is now scrubbed on the way out and
# rehydrated on the way back, transparently. Streaming (SSE) responses work.

See the CLI Reference for the full set of options and supported providers.

Node.js API: Scrubbing, Rehydrating, and Key Injection

import {
  scrub,
  rehydrate,
  setCachedEncryptionKey,
} from '@nanocollective/prompt-scrub';

// Optional: inject a key without relying on PROMPT_SCRUB_KEY / TTY prompting.
setCachedEncryptionKey(process.env.MY_APP_KEY ?? '');

const prompt = "My key is sk-12345";
const { scrubbedContent, sessionId } = scrub({ content: prompt });
console.log(scrubbedContent); // "My key is «Secret_1»"

// ... send to LLM ... get response "I see your key is «Secret_1»"

const { content } = rehydrate({
  content: "I see your key is «Secret_1»",
  sessionId,
});
console.log(content); // "I see your key is sk-12345"

Configuration

prompt-scrub reads an optional config file for extra rule packs, URL allowlisting, the active locale, a default confidence floor, session TTL, and encryption. Create one pre-filled with the default schema:

prompt-scrub init
# Created config file at /home/alice/.config/prompt-scrub/config.json

Print the configuration that is actually active — JSON on stdout, the path it came from on stderr:

prompt-scrub config show

Set locale (or pass --locale de-DE to scrub, inspect, or watch) to activate detectors from locale-specific rule packs. Built-in detectors are English-shaped and always run. A locale that activates no detector warns on stderr rather than silently scrubbing with English rules only.

Entries that do not match the schema are reported and exit non-zero, so config show doubles as a config check in scripts. See the CLI reference for the full schema and options.

Session Management & Garbage Collection

prompt-scrub maps sensitive data to placeholders and stores these mappings in session files (by default in ~/.config/prompt-scrub/sessions/).

To protect your privacy and save disk space, session files have a Time-to-Live (TTL). By default, sessions that are older than 7 days are automatically deleted when you run the scrub or sessions list commands. Active sessions (ones you interact with) will naturally have their TTL reset.

You can also manually prune expired sessions:

prompt-scrub sessions gc

To configure the TTL, add sessionTtlDays to your ~/.config/prompt-scrub/config.json:

{
  "sessionTtlDays": 14,
  "encryptionEnabled": true
}

Once encryptionEnabled is on, run prompt-scrub sessions encrypt (passing PROMPT_SCRUB_KEY) to rewrite any plaintext sessions already on disk.

Documentation

Full user guides and architecture details are in the docs/ directory:

• Getting Started
• Threat Model
• Architecture

Read the full whitepaper at docs.nanocollective.org.

Community

• Discord: Join the Nano Collective Discord
• Contributing: Read our Contributing Guide to get started.
• Issues: Check the GitHub Issues for planned work or to report bugs.