jev-guard (Clemens Schartmueller)

ALLOW means the call runs without interrupting you; ASK needs a human decision, where Antigravity prompts you while Claude Code and Codex block the call and tell the agent why; DENY blocks it. Examples from the README: git status is ALLOW, cat .env and rm -rf dist are ASK, and recursive deletion of your home directory is DENY.

From the project you want to protect, run npx --yes @clemensschartmueller/jev-guard@latest. It downloads the right binary for your platform, verifies its SHA-256 checksum, installs it to ~/.jevguard/bin, and adds hooks for the agents it detects (all supported agents if none are detected). The installer needs Node.js 18+.

The hooks themselves run a native binary, so no Node runtime is needed at decision time, but semantic evaluation needs a TypeSafe AI API key. Set TYPESAFE_API_KEY in the environment you start your agent from; the README shows the export line for Linux and macOS and the PowerShell form for Windows.

Credential files, workspace escapes, and trusted read-only commands are decided locally in under a millisecond by fast-path rules. Ambiguous or destructive calls are judged by TypeSafe's System One model, which also considers what you actually asked for, so a requested deletion is not questioned again.

Yes: audit mode logs what would have been blocked without interrupting anything. Running the installer again is safe, and existing hooks are preserved, so you can install alongside whatever agent configuration you already have.

It is Clemens Schartmueller's MIT-licensed open-source project, published to npm as @clemensschartmueller/jev-guard, and its semantic judgments go through TypeSafe AI's System One model using your API key. The README frames the tool as an independent guard that sits in the agent's hook pipeline.

The fast path only recognizes known risky patterns; anything else depends on a reachable TypeSafe API, and when it is unreachable jev-guard asks rather than allows. Catastrophic actions are never auto-allowed even when requested. The README publishes no benchmarks, so its speed claims are the project's own.