ZTDS Reference Server

零信任数据清洗参考MCP服务器,适用于Cursor和Claude Desktop。本地内存中的PII去标识化和确定性替代令牌化(RFC v1.0)。

文档

RFC v1.0 SPECIFICATION · IETF draft-sibiryakov-ztds-protocol-02 · OPEN CONSORTIUM

Client‑Side AI Data Sanitization

LIVE IN-RAM · 0 Sockets · 0.00 B Egress

Sample Payload:

0 entities tokenized

2. Sanitized Egress (Sent to LLM)

< 0.30 ms

Open Full Pipeline Inspector →

A. Cloud LLM Stream (Safe Tokens Only)

B. Restored Locally in Host RAM

[

IETF Internet-Draft

draft-sibiryakov-ztds-02 →

](https://ztds.ai/ietf/)[

Archival Zenodo DOI

10.5281/zenodo.22058770 →

](https://ztds.ai/whitepaper/)[

Median In-RAM Latency

0.28 ms · 0.00 B Egress →

](https://ztds.ai/benchmarks/)[

Cryptographic Attestation

Ed25519 Seal Validator →

](https://ztds.ai/verify/)

Architectural Comparison

Why Cloud AI Gateways & Destructive Masking Fail Enterprise DPO Review

Routing prompts through an external DLP API transmits unmasked personal data to another third-party cloud before redaction begins. Irreversible [REDACTED] masking breaks multi-entity reasoning and code generation. ZTDS resolves both inside local volatile memory.

ArchitectureExecution BoundaryLLM Reasoning FidelityGDPR Art. 28 SubprocessorsGDPR Art. 17 Vector EraseLatency
Cloud DLP / AI Proxy GatewayExternal vendor cloud serverVariable (depends on proxy state)+1 SaaS Processor & DPA RequiredRequires full index rebuild+180–600 ms WAN
Destructive Static MaskingApplication server regexBroken ([REDACTED] collapses coreference)None, if strictly localIrreversible data loss1–5 ms
ZTDS™ In-RAM Protocol (RFC v1.0)Client / Host Volatile RAM (0.00 B egress)Bijective tokens ([EMAIL_1]) + local restore0 Sanitization SubprocessorsDrop local token key; keep index< 0.8 ms In-RAM

GDPR Art. 28 & Art. 32 Safe Harbor Matrix → Read 1-Page CISO Legal Memorandum → Breach Liability & ROI Model →

Production Integration

Three High-Intent Enforcement Boundaries

Enforce identical RFC v1.0 surrogate tokenization across developer IDEs, headless backend RAG pipelines, and CI/CD release gates.

Plane 1 · IDE & Agent Stdio

Model Context Protocol (MCP) Security

Intercepts prompts, `.env` secrets, and SQL tool payloads inside local workstation RAM for Cursor, Claude Desktop, and Windsurf over JSON-RPC `stdio` pipes.

ztds-mcp / @privacyscrubber MCP Spec →

Plane 2 · Node.js, TS & Python

Pre-Embedding RAG & Headless SDK

Sanitizes customer documents before Pinecone, Qdrant, or PGVector indexing and LangChain/LlamaIndex LLM calls, neutralizing vector inversion and GDPR Art. 17 re-indexing.

@privacyscrubber/sdk SDK Docs →

Plane 3 · DevSecOps & Edge

CI/CD Static Gate & Browser Intake

Blocks unguarded LLM SDK calls in pull requests via npx ztds-audit --strict and issues offline-verifiable SHA-256 & Ed25519 compliance receipts.

GitHub Actions / CLI 10 Blueprints →

// npm install @privacyscrubber/sdk
import { ZTDSEngine } from '@privacyscrubber/sdk';

const engine = new ZTDSEngine({ license: process.env.ZTDS_LICENSE });
const { text: safePrompt, tokenMap } = await engine.sanitize(rawCustomerPrompt);

const llmReply = await openai.responses.create({ model: 'gpt-4.1', input: safePrompt });
const finalOutput = engine.restore(llmReply.output_text, tokenMap); // Restored in host RAM

Additional Enterprise Surfaces: Cloudflare Workers, Envoy WASM, Evidentiary Sovereignty & Regulatory Mappings ↓

OWASP LLM02, NIST AI RMF & EU AI Act Mappings → FRE 502 Evidentiary Sovereignty → 3-Tier Conformance Certification → SOC 2 & Trust Center →

Normative Specification

The Four ZTDS™ RFC v1.0 Invariants

Read Full RFC v1.0 Specification →

INVARIANT 01

Zero Pre-Sanitization Egress

Cleartext PII, PHI, and credentials never cross the client or host execution perimeter prior to tokenization.

INVARIANT 02

Deterministic Surrogates

Context-preserving tokens ([EMAIL_1]) maintain entity coreference while mapping tables stay in volatile RAM.

INVARIANT 03

Volatile Memory Isolation

Zero disk writes, zero third-party telemetry beacons, and immediate session zeroization upon completion.

INVARIANT 04

Zero Sanitization Subprocessors

Excludes external sanitization vendors from GDPR Art. 28 chains and supports GDPR Art. 32 technical safe harbor.

Apache 2.0 Open Reference

Reference CLI & Baseline MCP

Vendor-neutral RFC v1.0 test vectors, ztds-mcp baseline server, and npx ztds-audit static pipeline verifier for individual engineers and open-source projects.

Verified Registry → Case Studies

Pioneer Certified Implementation

PrivacyScrubber SDK, MCP & TEAMS

Production zero-network appliance engineered by BrandMeWeb with 30 Specialized Industry Profiles (HIPAA Safe Harbor, PCI-DSS, SWIFT/IBAN, Legal Privilege, Cloud IAM) and offline Ed25519 token licensing.

Commercial Tiers →

Turnkey Engineering

BrandMeWeb Enterprise Advisory

Asynchronous perimeter architecture audit, custom MCP/RAG enclave integration, SIG Lite & CAIQ vendor documentation, and formal ZTDS Verified™ attestation.

Scope SOW Workbench → BrandMeWeb ↗

Technical & Legal Reference

Frequently Asked Questions

How does ZTDS protect sensitive data in Model Context Protocol (MCP) for Cursor and Claude Desktop? ↓

The open reference MCP server (ztds-mcp) and certified @privacyscrubber/mcp-server execute locally on the developer workstation over stdio JSON-RPC pipes. Prompts, code diffs, and tool arguments are sanitized in volatile host RAM before network transmission. Cleartext secrets, internal IPs, and customer PII are replaced with deterministic surrogate tokens ([KEY_1], [EMAIL_1]) so LLM reasoning stays intact while 0.00 bytes of raw sensitive data cross the network.

How does ZTDS solve GDPR Article 17 Right to be Forgotten in RAG vector databases? ↓

When raw PII is embedded directly into Pinecone, Qdrant, or PGVector, honoring a GDPR Article 17 erasure request requires costly re-indexing. ZTDS tokenizes direct identifiers in host RAM prior to embedding generation. Vector indices store mathematical representations of de-identified surrogate tokens rather than raw personal identifiers, and purging the local token map severs re-identification without rebuilding the vector index.

How does ZTDS eliminate intermediary sanitization subprocessors under GDPR Article 28? ↓

Cloud DLP gateways and SaaS redaction proxies ingest cleartext personal data on external servers, acting as third-party processors requiring GDPR Article 28 Data Processing Agreements (DPAs).

Because ZTDS executes 100% in volatile client RAM with 0.00 bytes sent to any sanitization vendor, zero sanitization subprocessor enters the chain—supporting Technical Safe Harbor under GDPR Article 32 and Israel Amendment 13 Section 17.

What is the difference between open reference ztds-mcp and @privacyscrubber/mcp-server? ↓

ztds-mcp is the Apache 2.0 open-source reference server maintained by the ZTDS Consortium for baseline universal PII and secret redaction. @privacyscrubber/mcp-server (and @privacyscrubber/sdk) is the Pioneer Certified Commercial Implementation engineered by BrandMeWeb, adding 30 specialized industry profiles (HIPAA Safe Harbor PHI, PCI-DSS banking/IBAN, legal privilege, cloud IAM) and offline Ed25519 token licensing for air-gapped enterprise teams.

How can I independently verify zero external network egress in my own environment? ↓

ZTDS follows the Airplane Mode Verification Protocol: load this page or install npx ztds-audit --strict, disable Wi-Fi and Ethernet interfaces, and execute tokenization and de-tokenization with 100% functionality. Inspecting Chrome DevTools Network tab or packet captures confirms 0 outbound requests and 0.00 bytes of network egress.