Agentinel is a security layer designed for autonomous coding AI agents. It acts as a zero-cost, locally-run package guardrail, intercepting potentially malicious or hallucinated npm packages before they can be executed. This tool is crucial because AI agents, while powerful coders, can sometimes misremember or invent npm package names, leading to the installation of fake or harmful packages.

• 100% Local & Private: Operates entirely locally with zero cloud calls, ensuring data privacy.

• Ultra-low Latency: Scans packages with minimal delay, ensuring a smooth workflow.

• Extensive Threat Database: Utilizes a local OSV (Open Source Vulnerabilities) database and heuristic checks to identify threats.

• AI Self-Correction Loop: Feeds structured JSON feedback back into the AI agent's context upon blocking a package, enabling autonomous correction.

• Deep Tree Scanning: Scans the entire dependency tree of a package, not just the top-level one.

• Slopsquatting Protection: Employs heuristics based on package age, downloads, and publisher drift to catch zero-day threats.

• CI and CD Fallbacks: Can be integrated into CI/CD pipelines like GitHub Actions and GitLab CI to gate deployments.

Agentinel works by registering as a PreToolUse hook with the AI agent runtime. When an install command is proposed, Agentinel intercepts it, resolves the dependency tree locally, and scans each package against its threat database and heuristics. If a threat is detected, it halts the install (or warns in warn mode) and provides feedback to the AI for self-correction. Unlike commercial alternatives, Agentinel is free, private, and integrates directly at the AI hook level.