🛡️ SentinelMCP: Production-Grade LLM Security Guard

Secure your LLM workflows against Prompt Injections, Data Leaks, Credentials Exposure, and Phishing URLs — running natively inside Claude Desktop or any MCP-compatible environment.

SentinelMCP is an advanced, production-grade Model Context Protocol (MCP) server that acts as a secure firewall for Large Language Models. It analyzes inputs and outputs in real-time, masking sensitive data (PII) using Zod validation, serving dynamic audit stats as resources, and providing pre-configured system audit prompts.

📊 Security Impact & Efficacy Comparison

[!NOTE] Figures are based on SentinelMCP's internal test suites containing simulated threat payloads, jailbreaks, and PII datasets. Results may vary depending on LLM model prompts and system instructions.

🏗️ Architecture

graph LR
    A[User Input / Prompt] --> B[SentinelMCP Server]
    B --> C{Security Gateways}
    C --> D[Prompt Injection Scanner]
    C --> E[PII & Data Masker]
    C --> F[Secrets & Keys Detector]
    C --> G[URL Reputation Auditor]
    D & E & F & G --> H[Clean & Verified Payload]
    H --> I[Target LLM Engine]

⚡ Core Components

1. Tools (JSON-RPC Actions)

All inputs are validated using zod and automatically formatted to JSON Schema using zod-to-json-schema.

📊 2. Dynamic Resources

SentinelMCP exposes real-time session statistics and configuration details directly to the LLM Client:

• `ai-security://rules/active`: Active regular expressions, rule weights, and blacklisted command counts used by security checkers.
• `ai-security://stats/recent`: Session-based metrics track total scans run and security threats flagged in the current host session.

📝 3. Prompts (Templates)

Pre-packaged prompts to guide LLMs through systematic audit operations:

• `security-audit-helper`: Instantly guides the model through running full prompt injection, sensitive data, secrets leak, and URL trust audits on a given code block or prompt input.

🔌 Claude Desktop Integration

Link SentinelMCP directly to your local Claude Desktop application by adding it to your configurations (%APPDATA%/Claude/claude_desktop_config.json or ~/Library/Application Support/Claude/claude_desktop_config.json):

Option A: Quick Run with npx (Recommended)

Run SentinelMCP instantly without cloning the repo:

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

Option B: Local Production Build

If you cloned and built the project locally:

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

🎛 Honor Custom Safety Rules (security-rules.json)

Define your own keywords, system rules, or custom regex checks dynamically. Create a security-rules.json file in your workspace:

{
  "customPromptPatterns": [
    "custom-system-bypass-phrase",
    "my-test-jailbreak-trigger"
  ],
  "customSensitivePatterns": {
    "privateToken": "\\bsecret_token_[a-zA-Z0-9]{12}\\b"
  }
}

🛠️ Developer Setup & Test Coverage

Installation

npm install

Compile & Build

npm run build

Run Jest Unit Tests (100% Coverage passing)

npm run test

🛡️ OWASP LLM Top 10 Mapping

SentinelMCP directly addresses core vulnerabilities highlighted in the OWASP Top 10 for LLM Applications:

• LLM01: Prompt Injection ➔ Mitigated via scan_prompt_injection.
• LLM02: Insecure Output Handling ➔ Mitigated via audit_ai_output.
• LLM06: Sensitive Information Disclosure ➔ Mitigated via check_sensitive_data & detect_secrets.
• LLM10: Model Theft / Data Exfiltration ➔ Mitigated via check_url_safety.