🛡️ The Puli Guardian MCP

Puli Reviewer

is a proactive security and stability guardrail for your AI coding agents (Claude Code, Cursor, Antigravity, etc.). It prevents production breakages by confronting proposed code changes with a massive, real-time database of global incidents, edge cases, and security breaches.

Stop production errors before they happen.

Puli acts as a real-time bridge between your local development and global production insights. It doesn't just check syntax—it checks reality.

• Real-time Intelligence: Connects to a live DB of global incidents.
• Agent Native: Built specifically for the MCP ecosystem.
• IP Protection: Choose between BYOK or zero-retention cloud.

How do I install it?

🚀 How it Works

Whenver your coding agent attempts to modify your codebase,
Puli:

1. Analyzes the Context:
   Queries the agent to understand the intent of
   the change.

2. Cross-References Failures:
   Matches the code against a live-updated DB
   of real-world production errors and business-flow
   disruptions.

3. Stress Tests:
   Forces the agent to account for specific edge
   cases before the code is ever committed.
   
   

⚙️ Installation & Setup

1. Configure the MCP Server

Add the following to your coding agent's configuration (e.g., claude_desktop_config.json or your IDE's MCP settings):

{
  "mcpServers": {
    "puli-viewer": {
      "command": "uvx",
      "args": ["--from", "puli-mcp-server@latest", "puli-reviewer"],
      "env": {
        "BYO_OPENAI_API_KEY": "Optional* - <your-key-if-using-byok>",
        "BYO_GOOGLE_API_KEY": "Optional* - <your-key-if-using-byok>",
        "BYO_ANTHROPIC_API_KEY": "Optional* - <your-key-if-using-byok>"
      }
    }
  }
}

• Provide at lease one API key

2. Add Agent Rules

To ensure your agent utilizes the guardrail for every change, add our rules file to your project:

• For Cursor: Create a file at .cursor/rules/puli-reviewer.md and paste the content from our rules template here.

🛡️ Privacy & Security

We prioritize your IP security, Operations run locally on your station and no code is shared with our cloud.

We support two distinct working modes: