| name | vanish |
| description | Privacy scanner + opt-out orchestrator covering 210 data brokers, 30 AI training platforms, 8 face-search services, 13 third-party AI tools (incl. workforce-monitoring), 12 NCII-takedown destinations, plus LLM memorization probing, training-dataset membership checks, and AI history cleanup. Free agent-native alternative to DeleteMe / Optery / Incogni that also covers AI-era threats they don't. Audit-signed (HMAC-SHA256), local-first, triple-confirm safety gates for high-risk actions. |
| version | 0.3.0 |
| metadata | {"openclaw":{"requires":{"env":["VANISH_AUDIT_HMAC_KEY"],"bins":["node"]},"optionalEnv":[{"name":"VANISH_SECRET_MASTER_KEY","when":"when using the encrypted secret-store with --unlock-master-key (AES-GCM fallback path used only when Windows DPAPI unavailable)"},{"name":"OPENAI_API_KEY","when":"ONLY for `vanish llm-memory-check` against OpenAI. Not needed for any other subcommand. --dry-run bypass exists."},{"name":"ANTHROPIC_API_KEY","when":"ONLY for `vanish llm-memory-check` against Anthropic Claude. Same --dry-run bypass."},{"name":"<BROKER>_LIVE_ENDPOINT","when":"ONLY for `vanish b1-live` (experimental adapter). Overrides per-broker HTTP submission URL. Default: postman-echo.com test endpoint."},{"name":"<BROKER>_OFFICIAL_*_ENDPOINT","when":"ONLY for official broker live mode. Compliance-gated — requires termsAccepted, lawfulBasis, operatorId before use."}],"primaryEnv":"VANISH_AUDIT_HMAC_KEY","emoji":"🔍","homepage":"https://github.com/RAMBOXIE/vanish","os":["macos","linux","windows"],"always":false,"skillKey":"vanish"}} |
vanish
Privacy scanner + opt-out orchestrator. Free open-source alternative to DeleteMe ($129+/yr), Optery ($99+/yr), Incogni ($99+/yr) that also covers the AI-era threats they don't. Agent-native, audit-signed, local-first.
🔎 Clawhub compliance declaration
This skill:
- Never runs automatically.
always: false in frontmatter. No per-reply hook, no cron install, no scheduled task.
- Never sends notifications on its own.
--notify flag exists only as a dry-run placeholder (validated against none|telegram|email|signal) — no SMTP, Twilio, Signal, Telegram, SendGrid, or similar client library is imported or wired anywhere.
- Only writes files to declared paths. See "Filesystem access" section below for the exhaustive list.
- Only makes network calls to declared endpoints. See "Network access" section for every outbound target.
- Declares all environment variables it reads. See
optionalEnv in frontmatter for conditional requirements (e.g., OPENAI_API_KEY is only read when the user explicitly runs vanish llm-memory-check without --dry-run).
- Never installs system-level daemons — no
crontab entries, no launchctl services, no schtasks, no systemd units.
- Invokes a fixed allowlist of system binaries for browser-opening and clipboard ops (see "System binaries invoked").
Capabilities — 20 subcommands across 5 threat surfaces
Stability tiers: Vanish classifies every routed subcommand into one of four tiers — Core (8: stable, in hero), Specialist (4: stable, narrower scope), Labs (4: experimental / low-evidence / UX layers), and Internal helpers (4: plumbing). The canonical mapping with tier + evidence grade (A/B/C/D) per command lives in src/command-manifest.mjs; README's Capability matrix and CLI's vanish --help both render from it. The threat-surface grouping below is for narrative; cross-reference the manifest for the stability call.
🏢 Data brokers (original v0.2 scope)
scan [Core, evidence C] — Heuristic triage across 210 brokers. 0-100 score. Pure local computation, zero network calls. 10-second wall clock.opt-out [Core, evidence B] — Browser-assisted opt-out walkthrough for 58 brokers (opens browser, pre-fills clipboard, you solve captchas). Records HMAC-signed audit + 30-day verify follow-up.verify [Core, evidence A/B] — Re-check past opt-out submissions. Broker entries: HTTP liveness (A). AI-platform / face-service entries: reminder walkthrough with manual confirm (B). One-shot kinds (takedown, history) explicitly skipped.b1-live [Labs, evidence D] — ⚠️ Labs / experimental live HTTP submission adapter for 8 brokers. Captchas block real use. Defaults to postman-echo.com test endpoint; only Spokeo has a verifiable substitute path in MVP. Use opt-out for real work.cleanup [Internal helper], wizard [Labs, UX layer], queue [Internal helper], report [Core, evidence A], dashboard [Labs, UX layer] — supporting workflow commands (dry-run cleanup, 18-state conversational wizard, queue management, proof-report rendering, static dashboard builder). Plus b1-demo and dashboard:watch as Internal helpers.
🤖 AI training exposure (v0.3)
ai-scan [Core, evidence C] — Classify 30 LLM platforms (ChatGPT / Claude / Gemini / LinkedIn / Reddit / Cursor / etc.) as exposed / licensed / safe / action-needed. Zero network calls. Takes only platform names, no personal data.ai-opt-out [Core, evidence B] — Walkthrough + browser-assisted opt-out for 26 of the 30. Each tool has exact UI toggle name + step-by-step + tier overrides. 60-day reverify (AI platforms silently reset settings).
👤 Face-search (v0.3)
face-scan [Core, evidence C] — Directory of 8 face-recognition services (PimEyes, FaceCheck.ID, FindClone, Lenso, TinEye, Yandex, Google Lens, Clearview AI). Vanish never handles your photo — opens each service's own search page + prints walkthrough.face-opt-out [Core, evidence B] — Guided deletion requests including Clearview (CCPA/GDPR — LE-only DB you can't search). HMAC audit + 30-day (60 for Clearview) reverify.
🧠 Advanced AI-era checks (v0.3)
llm-memory-check [Labs, evidence D] — Probe OpenAI + Anthropic with 15 stalker-style prompts, detect verbatim leaks of user identifiers. Requires user's own API keys in env (optional: --dry-run uses a mock provider). ~$0.01/scan. A clean result does NOT prove safety — paraphrased knowledge slips through.dataset-check [Specialist, evidence B (Common Crawl) / C (others)] — Check if your URL is in Common Crawl (real CDX API query) / LAION / Pile / C4 / WebText / RedPajama / Dolma / FineWeb (walkthrough). Only outbound call: index.commoncrawl.org for CC lookup.clean-ai-history [Specialist, evidence B/C] — Discover where AI tools store conversation history on your disk; prints exact shell command to delete. Vanish never runs rm for you — you copy-paste. Covers 9 tools (Cursor / VS Code Copilot / Claude Desktop / ChatGPT Desktop + 5 web).third-party-ai [Specialist, evidence B] — Catalog + objection-letter generator for 22 AI tools others use on you (workplace meetings, HR/recruiting AI, medical scribes, workforce-monitoring agents). Jurisdiction-cited letters (GDPR Art 21/22/88, CCPA / AB-331, Illinois AIVIA + BIPA, NY Local Law 144 + Electronic Monitoring Act, HIPAA, German BetrVG §87). Includes --detect-installed to scan local machine for 8 commercial workforce-monitoring vendors (ActivTrak / Teramind / Hubstaff / Time Doctor / Insightful / Veriato / InterGuard / Viva Insights).
🛡️ NCII / leak takedown (v0.3)
takedown [Specialist, evidence B] — DMCA letters for 12 leak sites + StopNCII.org hash-registry walkthrough + Google intimate-imagery form + 4 legal templates (DMCA §512(c) / C&D / police report / civil pre-suit). Jurisdictions: US federal (DMCA / SHIELD Act / Take It Down Act 2025), EU GDPR, UK OSA, Canada §162.1, Australia OSA. Vanish never submits — drafts only.
Network access (all outbound targets)
Scan + offline subcommands (zero network): scan, ai-scan, face-scan, third-party-ai (without --detect-installed is offline too), takedown (drafts only), clean-ai-history (read-only path discovery), llm-memory-check --dry-run.
Subcommands with declared outbound HTTP:
| Subcommand | Endpoint | When | User consent |
|---|
verify | Broker profile URLs supplied by previous opt-out runs | HTTP liveness check on stored profileUrl | User added the URL via prior opt-out |
b1-live | postman-echo.com (default) or <BROKER>_LIVE_ENDPOINT | EXPERIMENTAL live submission | Explicit --live flag required |
dataset-check | index.commoncrawl.org CDX API | Only when --url + Common Crawl are selected | User supplies the URL |
llm-memory-check | api.openai.com, api.anthropic.com | Only without --dry-run | User supplies their own API key + specifies providers |
ai-opt-out, face-opt-out, face-scan, clean-ai-history, takedown | none — these only spawn the local OS "open URL" binary to launch the user's browser. No HTTP from the Node process. | — | User can pass --no-open |
Every outbound request sends a clear User-Agent: Vanish-Verify/1.0 (opt-out verification; https://github.com/RAMBOXIE/vanish) for broker verify, or the standard SDK UA for OpenAI/Anthropic.
Filesystem access
Writes
| Location | Used by | Data |
|---|
data/queue-state.json (default) or --state-file <path> | All subcommands producing audit | Follow-up queue + HMAC-signed audit log |
data/queue-state.json.lock | Queue mutations | Transient lock file (cleaned post-mutation; stale-lock detection has 30s TTL) |
data/*.json | dashboard (build-dashboard-data) | Denormalized queue state for the static dashboard HTML |
User-supplied --output <path> | scan, ai-scan, third-party-ai, takedown, dataset-check --json | Report/letter files |
os.tmpdir()/vanish-* | Tests only | Test isolation dirs (cleaned in finally blocks) |
| Per-secret file under secret-store directory | auth/secret-store | scrypt-KDF + AES-GCM (Windows: DPAPI preferred) |
Reads
- Catalog JSON files (read-only, bundled in repo):
src/adapters/brokers/config/broker-catalog.json, src/ai-scanner/ai-platforms-catalog.json, src/face-scanner/face-services-catalog.json, src/ai-history/history-catalog.json, src/dataset-check/datasets-catalog.json, src/third-party-ai/third-party-catalog.json, src/takedown/takedown-catalog.json, src/llm-memory/probe-catalog.json.
data/queue-state.json for audit/verify.- User-supplied paths via
--state-file, --output-json, --output-md, --share-card, --profile-url.
Scans (stat() only, read-only, never modifies)
clean-ai-history stats per-OS AI-tool cache paths (e.g., %APPDATA%\Cursor\User\History, ~/Library/Application Support/Cursor/) — reports size, prints shell command, never deletes.third-party-ai --detect-installed stats per-OS workforce-monitoring install paths (e.g., %PROGRAMFILES%\ActivTrak, /Applications/Teramind.app) — reports what's installed, never modifies.
Never writes outside: user-supplied paths, data/, os.tmpdir(). No fs.rename or fs.unlink calls outside the lock-file and session-temp lifecycle.
System binaries invoked
| Binary | Use | Subcommands | Safe? |
|---|
cmd /c start | Windows — open URL in default browser | opt-out, ai-opt-out, face-opt-out, face-scan, clean-ai-history, takedown | Yes — user-facing browser only |
open | macOS — open URL | same | Yes |
xdg-open | Linux — open URL | same | Yes |
clip | Windows clipboard write | ai-opt-out --clipboard | Yes — copies target setting name so user can Ctrl+F on the page |
pbcopy | macOS clipboard | same | Yes |
xclip / xsel | Linux clipboard | same | Yes (attempts both; silently no-ops if neither installed) |
All browser-open + clipboard invocations honor --no-open to skip in scripted/test mode.
Operating principles
- Manual trigger only — every subcommand requires explicit invocation.
always: false in frontmatter. No automation hooks.
- Triple YES confirmation for any high-risk action (3 separate wizard states, each needing exact
YES).
- Export decision (
yes / no) required before any delete.
- User-selected notification —
--notify is a dry-run placeholder; no actual sending.
- User owns legal judgment — Vanish provides workflow capability, not legal advice. Templates cite real law but are not legal representation.
- Credentials: minimum scope, shortest TTL, encrypted storage, post-task wipe. HMAC key required in production.
- Never submit destructive actions on the user's behalf — we open pages, show exact commands, record audits. The user clicks submit, runs
rm, sends the letter.
Safety gates enforced by code
| Gate | Enforcement |
|---|
| Manual trigger | --manual flag required for cleanup; missing flag returns blocked with clear nextActions |
| Triple confirmation | Three separate states each require exact YES; any mismatch blocks |
| Export decision | Must be yes or no before EXECUTE; ask without answer = hard block |
| HMAC signing | Every persisted state mutation signed with VANISH_AUDIT_HMAC_KEY |
| Credential lifetime | TTL enforced by AuthSession.validate({ minTtlSeconds }) |
| Secret storage | scrypt KDF + per-secret salt; Windows DPAPI preferred, AES-GCM fallback |
| Compliance block | b1-live official mode requires termsAccepted, lawfulBasis, operatorId |
| Destructive action gating | clean-ai-history only prints commands; takedown only drafts letters; face-scan/face-opt-out never upload photos |
Run examples
Scan (zero network, recommended first step)
vanish scan --name "John Doe" --email "j@example.com" --phone "+15550101"
vanish scan --name "..." --output-md ./my-report.md
AI training exposure
vanish ai-scan --linkedin --chatgpt --cursor
vanish ai-opt-out --chatgpt --linkedin
Face search
vanish face-scan --pimeyes --facecheck
vanish face-opt-out --pimeyes --clearview
Workforce monitoring
vanish third-party-ai --detect-installed
vanish third-party-ai --context workforce-monitoring \
--detect-installed \
--jurisdiction US-state-NY-EMA \
--company "Acme Corp" \
--output workforce-objection.md
NCII takedown
vanish takedown --stopncii
vanish takedown --dmca-letter --all-leak-sites \
--name "Your Name" --email "legal@you.com" \
--output dmca-package.md
Full interactive wizard
vanish wizard
Zero-install via npx
npx github:RAMBOXIE/vanish scan --name "..." --email "..."
Coverage summary
| Threat surface | Catalog size | Walkthrough count | Live-adapter count |
|---|
| Data brokers | 210 | 58 | 8 (experimental) |
| AI training platforms | 30 | 26 | 0 |
| Face-search services | 8 | 8 | 0 |
| Third-party AI tools (incl. workforce-monitoring) | 22 | — | 0 |
| NCII takedown destinations | 12 leak sites + 3 hash registries + 3 search engines | — | 0 |
| Training datasets | 8 | — | 1 (Common Crawl CDX) |
| AI history locations | 9 (4 local-app + 5 web) | — | 0 |
| LLM memorization providers | 2 (OpenAI + Anthropic) | — | API-based, user-key |
The three "coverage" columns are NOT the same thing. Triage ≠ walkthrough ≠ automated submission. See the capability matrix in README for the distinction.
Tests
334 tests across 26 files. Every commit runs 6-job matrix: Ubuntu/macOS/Windows × Node 20/22.
Coverage includes: safety gates · queue state persistence · retry/dead-letter routing · secret encryption (scrypt KDF) · HMAC audit signing with canonical JSON · timing-safe verification · wizard state transitions · scan scoring heuristic · 210-broker catalog registration · live HTTP round-trip against postman-echo · 30/60-day verify reverification · browser-assisted opt-out flows · share card rendering (broker + triple-threat v2) · AI training classification · face-scan service directory · LLM leak detection with mock fetch · Common Crawl CDX query with mock fetch · AI history cross-platform path expansion · third-party AI objection-letter rendering · workforce-monitoring detection with mock filesystem · takedown DMCA letter substitution · legal-jurisdiction clause dispatch across 10 jurisdictions.
Run: npm test
Non-goals (explicit scope boundaries)
Vanish does NOT:
- Run as a background service or scheduled task
- Send email, SMS, or push notifications
- Kill processes, terminate services, or alter system configuration
- Block phone-home traffic from monitoring tools
- Provide anti-detection or anti-forensics against monitoring software
- Auto-submit legal documents or send letters for you
- Auto-delete files from your disk (print commands; you execute)
- Upload your photos, emails, documents, or identity information anywhere
These are deliberate. Every capability is scoped to identification + documentation + evidence + drafted-request generation. The user is always the final actor.
Attribution
DeleteMe®, Optery®, and Incogni® are trademarks of their respective owners (Abine, Inc.; Optery, Inc.; Surfshark B.V.). Vanish is not affiliated with, endorsed by, or sponsored by any of these services. References exist solely for factual comparison (truthful comparative advertising permitted under US Lanham Act §43(a), EU Directive 2006/114/EC).
Pricing references ($99+/yr, $129+/yr) are entry-tier and approximate as of April 2026.
MIT-licensed — see LICENSE.
