Open-source plugin · Third-party
jev-sentinel
Pi, Claude Code, and Codex CLI guard that asks Jev whether a tool call is on-task, risky, or injected before it runs.
- Runnable
- Runnable as a Pi extension (git:github.com/harshwasan/jev-sentinel) or as a built hook for Claude Code and Codex CLI.
- Keys
- TYPESAFE_API_KEY is required. The project scrubs secret-looking strings before they reach Jev; that is pattern-based, not a guarantee.
- Platforms
- Pi coding agent, Claude Code hooks, Codex CLI hooks, Node.js
- Primitives
- choice, score, noul
pi install git:github.com/harshwasan/jev-sentinelPi has no built-in permission system. jev-sentinel puts a typed Jev judgment in front of tool calls, file reads, and replies: on-task versus injected, and a risk score that your host turns into allow, ask, or warn.
The same checks ship for Claude Code and Codex CLI through a stdin/stdout hook. Setup is documented in hooks/README.md. Decisions can be logged locally to ~/.jev-sentinel/decisions.jsonl.
This is a practical pattern for System One models: small questions, then code. It is not a replacement for sandboxing or secret management.
Caveats
- Third-party guard, not official TypeSafe software.
- A hook sees tool output after the tool already ran, unlike the Pi extension which can warn before the agent reads a file.
- Treat Jev probabilities as signals. Keep irreversible actions behind your existing approval rules.