Custodyn - Runtime Trust for AI Agents

Hi everyone 👋 I'm Atul, the creator of Custodyn. I spent years working in offensive security and noticed the same pattern every time AI agents became more capable: teams invested heavily in making agents smarter, but very little in controlling what those agents could actually do once connected to real systems. AI agents can now send emails, modify files, execute shell commands, call production APIs, and automate critical workflows. Once they're given real permissions, prompt engineering alone isn't enough—you need controls at execution time. That's why I built Custodyn. Custodyn sits between an AI agent and the systems it interacts with, evaluating every high-impact action against your policies before it executes. It can allow, block, or require human approval, while creating a tamper-evident audit trail for every decision. The SDK is open source (MIT), takes only a few minutes to integrate, and supports Python and JavaScript with integrations for popular AI frameworks. I'd genuinely appreciate your feedback—especially if you're building AI agents or agentic infrastructure. What controls do you think every production AI agent should have? Thanks for checking out Custodyn!