Ztds

ztds-mcp & ztds-audit — Zero-Trust Data Sanitization (ZTDS™ RFC v1.0) MCP Server & Invariant Auditor

This repository contains the official Model Context Protocol (MCP) Server implementation (`ztds-mcp`) and Codebase Invariant Auditor (`ztds-audit`) for the Zero-Trust Data Sanitization (ZTDS™ RFC v1.0) standard. Maintained by the ZTDS AI Consortium.

• MCP Server Source Code (in this repository): packages/ztds-mcp/index.js
• MCP Manifest & Server Card: packages/ztds-mcp/manifest.json | server-card.json | smithery.yaml
• MCP Server Test Suite: packages/ztds-mcp/test.js
• CLI Invariant Auditor: bin/ztds-audit.js

ZTDS™ governs in-memory data sanitization for AI agents, IDE assistants (Claude Desktop, Claude Code, Cursor, Windsurf, Zed), retrieval-augmented generation (RAG), and Large Language Model (LLM) pipelines. It executes 100% locally in volatile device memory (RAM) with zero external network calls, zero disk writes, and zero subprocessors.

---

1. MCP Server Implementation (ztds-mcp)

The reference MCP server (packages/ztds-mcp/index.js) implements the Model Context Protocol (2024-11-05) over stdio JSON-RPC 2.0 transport with zero external runtime dependencies. It intercepts sensitive cleartext (PII, PHI, API secrets, IBANs, credit cards) in client RAM prior to LLM dispatch, substitutes deterministic surrogate tokens ([EMAIL_TOKEN_1], [API_SECRET_TOKEN_1]), and restores authentic values locally in LLM responses.

1.1 Installation & MCP Client Configuration

Option A: 1-Click Automated Setup

npx ztds-mcp init

Option B: Claude Desktop (claude_desktop_config.json) & Cursor IDE (.cursor/mcp.json)

{

"mcpServers": { "ztds": { "command": "npx", "args": ["-y", "ztds-mcp"] } } }

Option C: Run Directly from Repository Source

{

"mcpServers": { "ztds": { "command": "node", "args": ["/absolute/path/to/ztds.ai/packages/ztds-mcp/index.js"] } } }

Option D: Claude Code CLI

claude mcp add ztds -- npx -y ztds-mcp

---

1.2 MCP Tools (tools/list & tools/call)

Implemented in packages/ztds-mcp/index.js:

Example JSON-RPC tools/call (ztds_sanitize -> ztds_restore)

Request (`ztds_sanitize`):

{

"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": { "name": "ztds_sanitize", "arguments": { "text": "Customer alice@acme.org reported failure with key sk-proj-demo-key.", "sessionId": "ide-session-01" } } }

Response:

{

"jsonrpc": "2.0", "id": 1, "result": { "content": [ { "type": "text", "text": "Customer [EMAIL_TOKEN_1] reported failure with key [API_SECRET_TOKEN_1].\n\n> [ZTDS Invariant 1 Attested] Zero external network egress.\n> Receipt: ZTDS-RECEIPT-8F4B2C19E0A3D712 | Execution: 42µs | Entities Masked: 2" } ], "isError": false } }

---

1.3 MCP Resources (resources/list & resources/read)

The MCP server exposes structured read-only protocol and telemetry resources via resources/list and resources/read:

---

1.4 MCP Prompts (prompts/list & prompts/get)

---

1.5 Programmatic / Node.js SDK Usage (packages/ztds-mcp/index.js)

The MCP server module exports its JSON-RPC dispatcher and in-memory sanitization primitives for direct programmatic integration and testing:

import {

handleMessage, sanitizeText, restoreText, auditText, resetSessionStore, TOOLS, PROMPTS, RESOURCES } from './packages/ztds-mcp/index.js';

// 1. Direct in-memory sanitization & restoration const scrubbed = sanitizeText('Send wire to DE89370400440532013000 for john@example.com', 'session-1'); console.log(scrubbed.sanitizedText); // "Send wire to [IBAN_TOKEN_1] for [EMAIL_TOKEN_1]"

const restored = restoreText(Confirmed transfer to ${scrubbed.sanitizedText}, 'session-1'); console.log(restored.restoredText);

// 2. Direct MCP JSON-RPC 2.0 protocol invocation const rpcResponse = await handleMessage({ jsonrpc: '2.0', id: 1, method: 'tools/list', params: {} });

// 3. Zeroize volatile session memory (Theorem 2) resetSessionStore('session-1');

Run the MCP server verification suite locally:

node packages/ztds-mcp/test.js

---

2. The 4 Fundamental Invariants (RFC v1.0)

Every AI pipeline protected by ztds-mcp and audited by ztds-audit is verified against the 4 core mathematical and architectural invariants:

1. Invariant 1: Zero External Egress Prior to Sanitization: Raw sensitive cleartext (PII, PHI, API secrets, private financial numbers) must never cross the local execution boundary prior to reversible surrogate tokenization.
2. Invariant 2: Deterministic Reversible Tokenization: Semantic, syntactic, and structural context must be preserved for LLM reasoning while de-identification lookup tables remain strictly in local host volatile memory.
3. Invariant 3: Cryptographic & In-Memory Isolation: Operations run in volatile heap or hardware-attested enclaves (WASM, Nitro Enclaves, C-ABI) with zero unencrypted disk spill and immediate memory zeroization.
4. Invariant 4: Continuous Subprocessor Exclusion: Local client-side execution eliminates third-party transmission, legally exempting deployments from GDPR Article 28 data processing agreements (Zero-DPA).

---

3. Codebase Invariant Auditor CLI (ztds-audit)

Run a zero-install security and invariant audit across your codebase or AI agent repository:

npx ztds-audit --dir ./src --strict

Command-Line Flags

Example Output

[ZTDS] ZTDS.ai In-Memory Codebase Auditor (RFC v1.0 Conformance)

---

Directory: /home/runner/work/ai-agent/src Files Scanned: 48 files in 18ms Audit Hash: sha256:7f83b1657ff1fc53b92dc18148a1d65dfc2d4b1fa3d677284addd200126d9069

---

[PASS] CONFORMANCE CONFIRMED: 0 INVARIANT VIOLATIONS DETECTED All scanned files comply with ZTDS Invariant 1 (Zero-Egress) and Invariant 3 (RAM isolation).

[LEGAL STATUS] Qualifies for GDPR Article 28 DPA Exemption (Zero-Subprocessor Chain) [PERFORMANCE] 0.00 Bytes Sensitive WAN Egress | Pure In-Memory Execution

Next Steps for Builders & Maintainers:

1. Include this audit hash in your pull request: https://ztds.ai/apply/
2. Embed your Verified Trust Badge in README.md to claim your registry backlink:

---

4. GitHub Actions CI/CD Integration

Enforce zero-trust data sanitization on every pull request by creating .github/workflows/ztds-audit.yml:

name: ZTDS Zero-Trust Compliance Audit

on: push: branches: [ main, master ] pull_request: branches: [ main, master ]

jobs: audit: name: Verify ZTDS RFC v1.0 Invariants runs-on: ubuntu-latest steps:

• name: Checkout Codebase uses: actions/checkout@v4

• name: Setup Node.js Runtime uses: actions/setup-node@v4 with: node-version: '20'

• name: Execute ZTDS Invariant Audit uses: moxno/ztds.ai@main with: dir: './src' strict: true

---

5. Empirical AI Safety Benchmark (Frontier LLMs)

Verify cleartext leakage prevention, bijective fidelity, and microsecond latency across OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, Google Gemini 2.0 Pro, and DeepSeek-V3:

# Execute 25 multi-domain enterprise scenarios across 4 frontier LLM families

npx ztds-bench

Generate formal Markdown evaluation report & Ed25519 cryptographic certificate

npx