Zentric Protocol MCP

ZENTRIC PROTOCOL

Prompt injection detection & PII anonymization with a signed audit trail — for LLM apps and AI agents.

*Every prompt, RAG chunk and tool output inspected before it reaches your model — deterministic CLEARED / ANONYMIZED / BLOCKED verdicts with a signed audit record (SHA-256 + UUID + UTC) per request.* *Detection tells you what happened. The audit record is what you show your auditor — GDPR Art. 30 evidence today, EU AI Act readiness tomorrow.*

→ Get API key (free, 10,000 req/mo) · Quickstart · Integrity Report v1.0 (PDF)

A real prompt injection attempt. Caught in <0.1ms. Your model never sees it.

---

Repository Scope & Commercial License

This repository exists for transparency and contribution — not as a deployable alternative to the hosted service.

Cloning this repository does not give you access to the Zentric processing service. The detection engine that inspects prompts, detects PII, and generates signed audit reports runs on Zentric's infrastructure and requires an active license.

Why publish the middleware?

Because trust is infrastructure. You should be able to verify how authentication works, how your API key is validated, and how subscription state is checked before your requests reach the engine. We believe in auditability at every layer — including our own enforcement code.

Contributions welcome

We accept contributions to the middleware, webhook handler, and Supabase schema. Open a PR or file an issue. For security-related contributions, see the Security section.

Getting access

---

What is Zentric Protocol?

Zentric Protocol is an infrastructure integrity layer for AI systems. It sits between your application and your LLM, examining every signal — prompts, responses, user inputs — and returning a cryptographically-signed verdict before execution continues.

It is not a filter. It does not guess. It applies deterministic rules across a standardized pipeline and returns a structured, auditable JSON report for every request.

Input Signal

│ ▼ ┌─────────────────────────────────────────┐ │ ZENTRIC PROTOCOL │ │ │ │ ┌─────────────┐ ┌─────────────────┐ │ │ │IntegrityGuard│→│ PrivacyGuard │ │ │ │ 22 injection │ │ 12 PII types │ │ │ │ signatures │ │ 7 languages │ │ │ └─────────────┘ └────────┬────────┘ │ │ ▼ │ │ ┌──────────────┐ │ │ │ ZentricReport│ │ │ │ UUID+SHA-256 │ │ │ │ GDPR Art.30 │ │ │ └──────────────┘ │ └─────────────────────────────────────────┘ │ ▼ Verdict + Certificate → Your System

---

Performance

Zentric uses deterministic signature matching — not an ML classifier. Every block is a known pattern match, which means 100% precision on known patterns and zero false positives: nothing is ever blocked unless it matches a catalogued signature. Verdicts are returned in sub-millisecond time (no model in the hot path), so the same input always produces the same verdict.

Any published metric is reproducible: run npm run benchmark (benchmarks/run.mjs) against the public deepset/prompt-injections dataset to verify the numbers yourself.

---

The Three Modules

01 · IntegrityGuard

Detects prompt injection, jailbreak attempts, and instruction overrides before they reach your LLM.

• 22 catalogued injection signatures
• 7 supported languages (EN, ES, FR, DE, PT, ZH, JA)
• Deterministic multilingual signature matching — no ML model in the verdict path
• Mean server-side processing: <0.1ms *(sub-millisecond; no model in the hot path)*

02 · PrivacyGuard

Identifies and anonymizes PII in prompts and responses. Regional standards treated as first-class entities.

• 12 PII entity types, format-validated (Luhn, IBAN mod-97, mod-11, NIF/NIE checksum): SSN, NIF, CPF, CURP, IBAN, SWIFT, passport, email, phone, and more
• Regional pattern recognition (EU, US, LATAM)
• Anonymization operators: redact, mask, tokenize, pseudonymize

03 · ZentricReport

Every request that passes through the protocol generates a signed, immutable audit record.

{

"report_id": "zp_01HXYZ...", "uuid": "f47ac10b-58cc-4372-a567-0e02b2c3d479", "timestamp_utc": "2026-05-14T22:00:00.000Z", "sha256": "e3b0c44298fc1c149afb...", "verdict": "CLEARED", "integrity": { "injection_detected": false, "signatures_matched": [], "confidence": null }, "privacy": { "pii_detected": true, "entities": [ { "type": "EMAIL", "action": "REDACTED", "position": [42, 61] } ] }, "compliance": { "audit_record": true, "ccpa": true, "eu_ai_act_s52": true }, "latency_ms": 0.05 }

---

API Reference

Authentication

curl -X POST https://api.zentricprotocol.com/v1/analyze \

-H "Authorization: Bearer zp_live_..." \ -H "Content-Type: application/json" \ -d '{ "input": "Your prompt or user input here", "modules": ["integrity", "privacy"], "options": { "anonymize": true, "language": "auto" } }'

Response

{

"status": "ok", "verdict": "CLEARED", "report": { ... }, "anonymized_input": "Your prompt or user input here", "latency_ms": 0.05 }

Verdict States

SDKs

---

Compliance Coverage

Zentric Protocol is designed from the ground up for regulated AI deployments.

---

Pricing

→ See plans · → Get API key · → Contact for Enterprise

---

Architecture Principles

Deterministic. The same input always produces the same verdict. No probabilistic black boxes in the critical path.

Stateless. The protocol does not store your data. Each request is processed and returned. The audit record is yours.

Composable. Deploy the full stack, a single guard, or wire only the audit layer into existing infrastructure.

Auditable. Every verdict is signed with SHA-256, timestamped in UTC, and assigned a UUID. Your compliance team will thank you.

---

For Agent Pipelines

Agent attacks don't arrive through the chat input. They arrive through tool call responses, RAG chunks, and memory retrievals — any external content that enters the prompt window. Your system prompt doesn't protect you here: it doesn't run until after the input is already parsed.

Wire Zentric at every ingestion point, not just on user messages:

• LLM input — user messages before they reach the model
• Tool output — external API responses before they re-enter the context window
• RAG retrieval — document chunks before they are assembled into the prompt
• Memory reads — stored context before it is injected into the next turn

One POST to /v1/analyze. The verdict comes back in sub-millisecond time. The agent continues or halts based on the result. Nothing else changes in your pipeline.

curl -X POST https://api.zentricprotocol.com/v1/analyze \

-H "Authorization: Bearer zp_live_..." \ -H "Content-Type: application/json" \ -d '{"input": " ", "modules": ["integrity", "privacy"]}'

---

MCP Server — Claude Desktop Integration

Zentric Protocol ships a native Model Context Protocol (MCP) server that integrates directly with Claude Desktop and any MCP-compatible agent runtime.

What it does

The MCP server exposes Zentric's detection engine as a native MCP tool. When wired into Claude Desktop, the agent automatically calls analyze_prompt before sending any input to the LLM — user messages, tool responses, RAG chunks, and memory retrievals are all checked.

MCP Tool exposed

analyze_prompt(text: string) -> ZentricReport

Returns: verdict (CLEARED / BLOCKED), risk_score, matched signatures, pii_entities, report_hash (SHA-256), latency_ms.

Install via npm

npx zentric-protocol-mcp

Claude Desktop configuration

Add to your claude_desktop_config.json:

{

"mcpServers": { "zentric-protocol": { "command": "npx", "args": ["zentric-protocol-mcp"], "env": { "ZENTRIC_API_KEY": "your_api_key" } } } }

Get your API key at zentricprotocol.com/quickstart — free tier is 10,000 r