Kastra Introduces Runtime Authorization for AI Coding Agents

Introduction to Kastra

Kastra is designed for developers using AI coding agents like Claude Code, Codex, Cursor, and OpenClaw. It was created to address potential risks associated with AI agent operations, following a near-miss incident involving a destructive action on a production database.

Functionality of Kastra

Kastra intercepts tool calls from AI agents and evaluates them against predefined policies before execution. This allows users to create policies in plain English through a web app, and decisions can be categorized as allow, hold, or deny. An immutable audit trail captures all decisions made by the system.

Policy Packs and Risk Scans

The tool includes policy packs that cover common high-risk scenarios, enabling enhanced security. Additionally, Kastra features a scan command called Recon, which analyzes the local session history of agents to identify past risky actions. This feature helps developers understand the risk profile of their AI agents.

Implementation and Usage

The setup for Kastra is straightforward, with command lines for installation and execution provided. The tool scans for risky actions taken by agents, highlighting concerns such as unauthorized database accesses or unintended file modifications while ensuring that sensitive information remains local.

Conclusion

Kastra aims to empower developers to use AI agents more safely by enforcing deterministic policies on their actions. This will facilitate greater trust in AI operations while minimizing the risk of harmful activities.