Safe operations pass through instantly. Risky operations require human approval via dashboard or Telegram. Dangerous operations are blocked completely

I built Latch to address the growing security risks of AI agents accessing critical systems. There have been 1,800+ exposed agent gateways discovered in the wild and recent security audits showing multiple vulnerabilities in agent frameworks, so this was motivated by the clear urgent need for better controls.

Implementation is simple: a CLI wrapper around MCP servers that takes a few minutes to set up. You can do it through your dashboard or with quick command (npx @latchhq/cli wrap)

Key features:

Policy engine using natural language rules, or create rule-based policies by tool name, action class, or domain.

Audit logging

Telegram integration for mobile approvals

The project addresses what security researchers call the "lethal trifecta" - when AI agents have access to private data, process untrusted content, and can communicate externally. This creates a new attack vector traditional security tools don't address.

I think the space is wide open and looking for contributors!

GitHub: https://github.com/latchagent/latch Docs: https://latch.mintlify.app