jev-axi adds Jev safety checks to Claude Code
jev-axi plugs Jev into Claude Code to block risky Bash commands, leaked secrets, and prompt injection, falling back to normal permissions on timeout.
YOUR CLAUDE CODE AGENT DOESN'T NEED MORE POWER. IT NEEDS SOMETHING THAT SAYS NO. Jev is a small model that never writes a word. It only judges: yes/no, pick one, score. About half a second per call. jev-axi plugs it into Claude Code as a second opinion. 5-minute setup: → npm install -g jev-axi → jev-axi config set apiKey YOUR_KEY (key from http://console.typesafe.ai) → jev-axi setup safety --project risky Bash commands get checked before they run. download-and-run, exfiltration, destructive stuff → blocked. borderline → it asks you. tests, builds and read-only commands are decided locally and never leave your machine → jev-axi setup git-hooks a commit with a private key or API token in it → stopped. that scan runs locally too → jev-axi setup supervise --project agent looping, drifting or quitting early → it gets a warning in its context. warn-only by default → jev-axi guard < email.txt hidden "ignore previous instructions" in anything your agent reads → flagged, exit code 3, pipeline stops bonus: pipe a failing build log into jev-axi triage → it points at the root-cause line the part I like: it never auto-approves. if Jev times out or errors, Claude Code's normal permission flow takes over. honest note: guard sends the text to the API. don't pipe secrets into it. cost: the CLI's default estimate is $0.042 per 1M input tokens. one check = fractions of a cent.
This post quotes another post: x.com/thegreatest_sv/status/2102802378378621355