laya-guard

laya-guard: A Free Local Guardrail for AI Coding Agents

laya-guard is a robust and efficient guardrail system designed to evaluate every tool call made by AI coding agents locally. This approach ensures that all actions, including shell commands and data transactions, are vetted in real-time and without the need for costly per-decision API calls. Utilizing a compact multilingual model known as Laya (~322M parameters), laya-guard performs swift regex prefiltering to detect known dangerous patterns and provides judgment on any other action, categorizing them into three clear responses: allow, review, or block.

Key Features

• Local Execution: Every decision is made on-device, ensuring privacy and security without leaving sensitive data unprotected.
• Quick Response: Achieves judgment on all calls in approximately 0.1 seconds, providing immediate feedback for developers.
• Multiple Language Support: The multilingual capabilities of the Laya model accommodate diverse coding environments.

Dashboard Overview

laya-guard features a user-friendly dashboard that displays vital statistics about the activities being monitored:

• Judgment Statistics: View a summary of call outcomes including counts of allowed, reviewed, and blocked actions.
• Call History: Browse through recent calls with detailed insights into their status (allowed, reviewed, or blocked) and linked actions.
• Agent Monitoring: Automatically detects coding agents on the machine and indicates which have active hooks installed.

How It Works

When an AI coding agent attempts to execute a command:

1. The command passes through a installed hook that forwards it to the laya-guard server.
2. The server assesses the call using established regex rules and the Laya model.
3. The decision is returned: allowed actions proceed, reviews prompt human confirmation, and blocked commands are terminated.

Supported Agents

laya-guard includes seamless integrations with various coding agents, allowing easy installation and configuration. The supported agents include:

• Claude Code
• Codex CLI
• Cursor
• Gemini CLI
• Qwen Code
• Windsurf
• Copilot CLI
• And more...

Use Cases

laya-guard's flexible design supports a variety of applications:

• Guarding Code Execution: Implement safety measures for AI-driven coding assistants seamlessly.
• Auditing Commands: Run batch command checks to ensure compliance with security protocols.
• Integration with HTTP Frameworks: Protect tool execution in workflows powered by APIs when running custom automation.

With its emphasis on local processing and user-centric design, laya-guard provides a practical solution to enhance the security and reliability of AI coding agents. The project is actively maintained and encourages contributions and improvements from the community.