Kastra
A permission gate that sits in front of your AI coding agent and blocks it from running a dangerous command or deleting a file before it happens — instead of hoping the agent behaves.
🔗 Visit Kastra
Description
Letting an AI agent run shell commands, hit APIs, or query a database on its own is powerful and a little terrifying — one bad instruction and it can do real damage before anyone notices. Kastra is built to close that gap: it sits between the agent and the outside world, checking every action against a policy in under a millisecond, and can revoke access instantly if something looks wrong.
Kastra is a runtime authorization layer for agents built on Claude, Cursor, Codex and OpenClaw, providing policy decision points, signed append-only audit trails, and both cloud and self-hosted (including air-gapped) deployment. It supports SDKs across TypeScript, Python, Go, Rust, Java and Swift, and controls shell commands, API requests, database queries and browser actions specifically — not just a generic firewall. It's SOC 2 Type II, ISO 27001, HIPAA and GDPR compliant, aimed squarely at teams running agents in regulated or production environments.
💬 Our review
The short version: if you're letting AI coding agents touch production systems and haven't thought about what stops them from doing something catastrophic, Kastra is the missing safety layer, not an optional nice-to-have.
The sub-millisecond policy decision claim (p99 <0.8ms) is the right design goal — a permission layer that adds noticeable latency to every agent action would get bypassed or disabled in practice, so speed here isn't a vanity metric. Compliance certifications (SOC 2, HIPAA) matter because this is exactly the kind of tool a security or compliance team needs to sign off on before agents touch anything real. The free tier (50K decisions/month) is generous enough to actually evaluate it on a real project, not just a toy demo. The honest caveat: this is a young, unproven category — "agent governance" barely existed a year ago, so there's no long track record, and pricing per developer seat ($19.99–$49.99) can add up fast for larger teams running many agents in parallel.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Gratuit (50K décisions/mois) ; Pro 19,99$/dev/mois (1M décisions) ; Team 49,99$/dev/mois (10M décisions) ; Enterprise sur devis
Pros
Décisions sub-milliseconde, n'ajoute pas de latence perceptible
SOC 2/ISO 27001/HIPAA/GDPR
SDK multi-langages
Déploiement air-gap possible
Cons
Catégorie très jeune, peu de recul terrain
Prix par développeur peut grimper vite en équipe nombreuse
Pas de clients nommés publiquement