AI-runtime-guard – Policy enforcement layer for MCP AI agents

Name: AI-runtime-guard – Policy enforcement layer for MCP AI agents
Availability: InStock
Author: JimmyRacheta

by JimmyRacheta·Feb 25, 2026·2 points·2 comments

Visit Project View on HN

AI Analysis

●●●BangerSolve My ProblemZero to One

Policy enforcement layer stops AI agents from deleting files or leaking credentials—no prompt retraining needed.

Strengths

•Solves real, urgent problem: AI agents with shell access currently have no execution guardrails between agent decision and system impact
•MCP-agnostic design integrates with any stdio-compatible client without modifying agent or workflow
•Practical features (auto-backup, blast-radius simulation, audit trail, approval GUI) ship together as a product, not just blocks

Weaknesses

•Early access stage with unclear production stability; web GUI and advanced policy tiers incomplete per author description
•Relies on policy file maintenance; permissive defaults mean user must actively configure to get real protection

Post Description

I built this after realizing that AI agents with filesystem and shell access can delete files, leak credentials, or execute destructive commands — and there's no enforcement layer stopping them at the execution level.

ai-runtime-guard is an MCP server that sits between your AI agent and your system. It enforces a policy layer before any file or shell action takes effect. No retraining, no prompt engineering, no changes to your agent or workflow.

Your agent can say anything. It can only do what policy allows.

What it does: - Blocks dangerous commands (rm -rf, dd, shutdown, privilege escalation) before execution - Gates risky commands behind human approval via a web GUI - Simulates blast radius for wildcard operations before they run - Creates automatic backups before destructive actions - Full audit trail of everything the agent does

Works with Claude Desktop, Cursor, Codex, and any stdio MCP-compatible client. Default profile is basic protection out of the box — advanced tiers are opt-in.

Validated on macOS Apple Silicon. Linux expected to work, formal validation coming in v1.1.

Would love feedback from anyone running AI agents with filesystem access.