Kastra

An authorization layer that intercepts every AI action—prompts, tool calls, shell commands, and API requests—and checks them against policy before execution, solving the problem of ungoverned AI actions in production. It sells to developers, engineering teams, and security teams at companies of all sizes, from startups to enterprises, who build or deploy AI agents and coding assistants. Unlike monitoring tools that log actions after the fact, Kastra decides allow or deny in sub-millisecond before the action runs, and can be deployed cloud, self-hosted, or air-gapped.

Key features

• Runtime authorization for every AI action
• Attribute-based policy language (DSL)
• Post-inference validation of model outputs
• Local enforcement on developer laptops (Edge)
• Autonomous agent controls with human approval
• Signed, append-only audit trail
• History scanning with policy drafting (Recon)