Grith
OS-level security supervisor that intercepts and scores what AI coding agents can do on Linux.
🔗 Visit Grith
Description
Letting an AI coding agent run commands on your machine is convenient right up until it does something you didn't expect, deletes a file, reads a secret, or reaches out to a server it shouldn't. Grith sits underneath the AI agent at the operating-system level and checks every action before it happens.
Grith is an OS-level security supervisor for AI coding agents, including Claude, Codex, Aider, Cursor, Cline, and Copilot, that intercepts Linux system calls, scores each action against 18 built-in security filters covering secret scanning, sensitive-path detection, egress policy, destructive-operation checks, and taint tracking, and automatically allows low-risk actions, queues ambiguous ones for human review, and denies high-risk ones, all through a single wrapper command with no agent-specific integration required. It keeps a hash-chained audit log and can run fully offline. MPL-2.0 licensed, freemium.
💬 Our review
The short version: Grith is one of the more serious attempts at actually sandboxing what AI coding agents can do at the operating-system level, not just at the prompt level, a sensible free layer of protection if you're letting agents run with real system access.
Broader sandboxing tools like Firejail, gVisor, or plain seccomp and SELinux policies can restrict what a process does, but none of them are purpose-built for the specific risk patterns of AI coding agents, accidental secret leaks, unexpected destructive git or file operations, agents making network calls they shouldn't. Grith's 18 pre-built filters and per-agent profiles save you from hand-writing those policies yourself. The free tier runs fully offline with no telemetry, which matters for security-conscious teams; paid tiers add license validation and analytics sync for larger deployments. At only a handful of stars it's early, but the problem it addresses, unsupervised AI agents with shell access, is only getting more common, so it's worth evaluating now rather than after an incident.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Offre gratuite (hors-ligne, sans télémétrie) ; les offres payantes ajoutent validation de licence et synchronisation analytique
Pros
18 filtres de sécurité prêts à l'emploi spécifiques aux risques des agents IA
Fonctionne avec Claude, Codex, Aider, Cursor, Cline, Copilot d'emblée
Journal d'audit chaîné par hash pour preuve de conformité
Peut fonctionner entièrement hors-ligne sans télémétrie
Offre gratuite disponible
Cons
Linux uniquement
Projet jeune (4 stars)
Les outils de sandboxing génériques (Firejail, gVisor) sont plus établis bien que moins spécifiques aux agents