CloakClaw — Always-On Privacy Proxy

Prerequisites

CODEBLOCK0

Verify: cloakclaw --version should return 0.1.2 or higher.

How It Works

1. 1. Detect: When user sends a document or sensitive text, auto-detect the document type
2. Cloak: Run the cloaking engine (regex + local LLM) to replace sensitive entities with realistic fakes
3. Send: Forward cloaked text to the cloud LLM for analysis
4. Decloak: When response arrives, reverse all replacements to restore originals
5. Deliver: Send the restored response to user with a privacy footer

Entity Types (24)

Category	Types
Identity	People, Companies, Passports, Drivers License
Contact

Emails, Phones, Addresses | | Financial | Dollars, Percentages, Accounts, Banks, SSNs | | Legal | Case Numbers, Jurisdictions | | Tech | IP Addresses, MAC Addresses, Passwords/Secrets, API Keys, URLs | | Other | Crypto Wallets, GPS Coordinates, VIN Numbers, Medical IDs, Dates |

Profiles (6)

• - general — all 24 types (default for unknown documents)
• legal — contracts, NDAs, filings (10 types)
• financial — bank statements, P&L, investor docs (11 types)
• email — correspondence (10 types)
• code — .env files, configs, infra docs (9 types)
• medical — HIPAA-adjacent records (11 types)

Auto-Detection Rules

Always cloak (document attached):

• - PDF, TXT, MD, CSV, JSON, YAML, code files → auto-detect profile from content
• Legal keywords (agreement, contract, whereas, hereby) → legal profile
• Financial keywords (revenue, P&L, balance sheet, quarterly) → financial profile
• Code files (.env, .yaml, .json with secrets) → code profile
• Default for unrecognized → general profile

Always cloak (sensitive content in text):

• - Contains SSN patterns (###-##-####)
• Contains dollar amounts > $1,000
• Contains multiple proper names + company names
• Contains IP addresses, API keys, or passwords
• User explicitly says "cloak", "private", "redact", or "protect"

Skip cloaking:

• - Simple questions with no sensitive data
• User says "raw", "uncloak", "no cloak", or "cloakclaw off"

Execution Flow

Step 1: Cloak the document

CODEBLOCK1

Output JSON:
CODEBLOCK2

Or use the CLI directly:
CODEBLOCK3

Step 2: Send cloaked text to cloud LLM

Use the cloaked text as the document content. The user's question stays unchanged — only the document data is cloaked.

Step 3: Decloak the response

CODEBLOCK4

Or CLI:
CODEBLOCK5

Output: restored text with original entities.

Step 4: Deliver with privacy footer

Append to the response:
CODEBLOCK6

User Commands

• - cloakclaw off — disable auto-cloaking for this session
• INLINECODE7 — re-enable auto-cloaking
• INLINECODE8 — show current settings and recent sessions
• INLINECODE9 — show what was cloaked

Configuration

Config at ~/.cloakclaw/config.yaml:

CODEBLOCK7

Recommended Models by RAM

RAM	Model	Quality
8GB	qwen2.5:3b	Basic (regex does most work)
16GB

qwen2.5:7b | Good |

| 32GB+ | qwen2.5:32b | Very good | | 64GB+ | qwen2.5:72b | Excellent |

Security

• - AES-256-GCM encrypted mapping database
• Optional password protection (cloakclaw password set)
• Auto-expiry: sessions purged after 7 days
• Zero telemetry, zero cloud dependency for cloaking
• All processing runs locally

⚠️ Disclaimer

CloakClaw is NOT HIPAA, GDPR, SOC 2, PCI-DSS, or CCPA compliant. It is a best-effort privacy tool. Users are responsible for reviewing cloaked output before sharing.