ZTDS Data Sanitizer
Serveur MCP de référence pour l'assainissement des données Zero-Trust, destiné à Cursor et Claude Desktop. Désidentification locale des PII en mémoire et tokenisation déterministe par substituts (RFC v1.0, Apache-2.0).
Documentation
ZTDS MCP Server (ztds-mcp)
Local Security Firewall for AI Agents & IDEs (Cursor, Claude Code, Windsurf)
Open-source reference implementation of the Zero-Trust Data Sanitization (ZTDS) protocol for the Model Context Protocol (MCP) ecosystem. Conforms to the open architectural specification RFC v1.0 and IETF Standards Track sibiryakov-ztds-protocol.
Runs 100% locally in device volatile memory (RAM) with zero network calls, zero external subprocessors, zero disk writes, and zero telemetry.
1-Click Quickstart (Automated Setup)
Auto-configure your installed AI IDEs and clients with a single command:
npx ztds-mcp init
The configurator automatically:
- Detects installed clients: Cursor IDE, Claude Desktop, and Windsurf.
- Creates or safely merges the
ztdsserver block into their local configuration files. - Executes an in-memory cryptographic self-test to verify zero-leakage tokenization.
To verify existing client configurations:
npx ztds-mcp status
Docker Container Quickstart (Isolated Stdio Runtime)
Run the containerized ztds-mcp image (linux/amd64 and linux/arm64) with zero host filesystem access:
docker pull brandmeweb/ztds-mcp:latest
Configure in claude_desktop_config.json or .cursor/mcp.json:
{
"mcpServers": {
"ztds": {
"command": "docker",
"args": ["run", "-i", "--rm", "brandmeweb/ztds-mcp:latest"]
}
}
}
Standalone CLI Pre-Execution Auditor
Audit sensitive files, configuration directories, or CI/CD pipelines for credential leakage prior to AI agent ingestion:
# Scan a single sensitive file
npx ztds-mcp audit .env
# Scan an entire repository/source directory
npx ztds-mcp audit ./src
# Machine-readable JSON output for CI/CD gates (exits 1 on violation)
npx ztds-mcp audit --json
Outputs line-by-line category breakdowns, risk assessment, and cryptographic SHA-256 attestation receipts.
Why AI IDEs Leak Data & How ZTDS Stops It
Autonomous agents in modern IDEs (Cursor, Claude Code, Windsurf) routinely index repository files, including .env secrets, database credentials, production logs, customer emails, and API keys. When an agent crafts a prompt or calls a tool, these sensitive strings are transmitted in cleartext across the internet to frontier LLM APIs.
ZTDS MCP acts as an in-memory local security firewall:
- Intercepts: Tools and prompt text are evaluated inside client RAM prior to transmission.
- De-identifies: Real credentials and PII are replaced with bijective, context-preserving synthetic tokens (
[API_SECRET_TOKEN_1],[EMAIL_TOKEN_1]). - Restores: When the model returns code or instructions containing synthetic tokens, the local MCP server re-maps the cleartext back into the IDE response.
- Zero Egress: Real private keys and customer data never touch the network unmasked.
The 4 Core Protocol Invariants
- Invariant 1: Zero External Egress Prior to Sanitization
Cleartext PII, PHI, and credentials never cross the local execution boundary unmasked. - Invariant 2: Deterministic Context-Preserving Reversible Tokenization
Sensitive values are replaced by synthetic tokens maintaining syntactic and semantic context for LLMs. - Invariant 3: Verifiable Ephemeral RAM Isolation (Theorem 2 Zeroization)
Mapping tables exist strictly in volatile memory and are zeroized upon session termination. - Invariant 4: Subprocessor Chain Exclusion
Operates strictly as a local computational utility under GDPR Recital 26, rendering Data Processing Agreements (DPAs) unnecessary.
Manual Client Configuration
If you prefer configuring your clients manually without npx ztds-mcp init:
1. Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"ztds": {
"command": "npx",
"args": ["-y", "ztds-mcp"]
}
}
}
Config file locations:
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%\Claude\claude_desktop_config.json - Linux:
~/.config/Claude/claude_desktop_config.json
2. Cursor IDE
Add to .cursor/mcp.json in your workspace root:
{
"mcpServers": {
"ztds": {
"command": "npx",
"args": ["-y", "ztds-mcp"]
}
}
}
3. Windsurf / Codeium
Add to ~/.codeium/windsurf/mcp_config.json:
{
"mcpServers": {
"ztds": {
"command": "npx",
"args": ["-y", "ztds-mcp"]
}
}
}
Available MCP Tools, Resources & Prompts
MCP Tools (tools/list & tools/call)
| Tool | Input Schema | Description |
|---|---|---|
ztds_sanitize | { text: string, sessionId?: string } | Masks sensitive PII/credentials with deterministic surrogate tokens prior to LLM transmission. |
ztds_restore | { text: string, sessionId?: string } | Restores original cleartext from volatile in-memory mapping into LLM output. |
ztds_audit | { text: string } | Scans text for exposed credentials and PII, returning SHA-256 integrity receipt and risk score. |
ztds_info | {} | Retrieves RFC v1.0 standard details, academic citations, and enterprise documentation. |
ztds_reset_session | { sessionId?: string } | Purges and zeroizes all volatile session token mappings (Theorem 2). |
MCP Resources (resources/list & resources/read)
| Resource URI | MIME Type | Description |
|---|---|---|
ztds://spec/rfc-v1 | application/json | Formal ZTDS RFC v1.0 / IETF draft-sibiryakov-ztds-protocol specification and 4 core invariants. |
ztds://session/telemetry | application/json | Real-time volatile RAM session count and zero-egress attestation status. |
MCP Prompts (prompts/list & prompts/get)
| Prompt | Arguments | Description |
|---|---|---|
ztds_guard | task (string, required) | Activates the ZTDS Zero-Trust Data Sanitization guardrail in Claude or Cursor. |
Compliance Frameworks & Architectural Reference
Explore the vendor-neutral specification, security boundaries, and regulatory blueprints on ztds.ai:
- MCP In-RAM Security & Tool-Call Boundary Architecture
- ZTDS Headless Developer SDK & RAG Integration Guide
- ZTDS Protocol Specification RFC v1.0
- Enterprise Regulatory Compliance & Zero-Subprocessor Matrix
- Zero-Trust Network Egress Scanner & CLI Auditor
- Compliance Implementation Blueprints Hub
- MCP Stdio PII Sanitization & Tool Poisoning Mitigation
- HIPAA Safe Harbor De-Identification for LLM Prompts
- GDPR Articles 25 & 32 Technical Safe Harbor for AI Prompts
- EU AI Act Article 10 Data Governance & Prompt Hygiene
- NIST AI RMF Generative AI Profile PII Controls
- OWASP LLM02 Sensitive Information Disclosure Defense
- RAG Pipeline Client-Side PII Sanitization
- Client-Side RAM Sanitization vs. Cloud Proxy DLP
Certified Commercial Reference Implementation (PrivacyScrubber)
ZTDS is an open, vendor-neutral standard. The open-source ztds-mcp package serves as the reference implementation covering universal baseline entities (Email, Phone, SSN, Credit Cards, IPv4, IBAN, API Secrets).
For production enterprise workloads requiring continuous regulatory compliance and air-gapped guarantees, PrivacyScrubber is the pioneer certified implementation offering:
- 30 Specialized Industry Profiles: HIPAA PHI (18 identifiers), PCI-DSS 4.0, GLBA Financial, SEC 17a-4, CJIS Law Enforcement, FERPA Student Records, Defense ITAR/CMMC, FRE-502 Legal Work-Product.
- Agentic Guard Automation: Autonomous zero-trust tool wrappers (
guard_exec,guard_read_file,guard_apply_patch). - Air-Gapped Node Licensing: Disconnected Ed25519 cryptographic tokens without cloud telemetry.
- Headless SDK: Backend RAG pipeline redaction for Node.js / TypeScript / Python (
@privacyscrubber/sdk). - SOC 2 Type II Evidence Binder: Automated audit artifacts (
ztds-evidence-binder.json) for Drata, Vanta, and AuditBoard.
Deploy the commercial implementation:
# Production MCP Server
npm install -g @privacyscrubber/mcp-server
# Headless Backend SDK
npm install @privacyscrubber/sdk
Website: https://privacyscrubber.com
Commercial integration references:
- Production MCP Server & Tool-Call Boundary Redaction
- Headless Node.js, TypeScript & Python RAG SDK
- HIPAA Safe Harbor AI Scrubber for Healthcare
Verification & Self-Test
To run the offline test suite:
node test.js
Conforms to standard JSON-RPC 2.0 stdio protocol. Zero runtime dependencies.
License
Apache-2.0. Maintained by the ZTDS AI Consortium (Working Group WG-1).
Website: https://ztds.ai
Lead Architect: Ilya Sibiryakov (Author of ZTDS RFC v1.0, IETF Standards Track sibiryakov-ztds-protocol, U.S. & Int'l Patents Pending).