Claude Code safety hook uses a local small model plus Jev for ~600 nightly allow/deny calls
The author hands allow/ask/deny decisions to a local 1.5GB model at about 18ms and zero token cost, routes fuzzy cases to Jev or a human, and keeps commands on the local machine with only a config change to switch.
jev + claude code... i honestly don't get why every agent setup isn't running this yet. my safety hook makes ~600 decisions a night. now a 1.5 GB model makes them on my mac in ~18 ms each, $0 in tokens, and the commands never leave the disk. → plain code rules catch the obvious commands before any model is called → narrow allow / ask / deny calls go to the local model → fuzzy ones get routed to jev or to me → moving a question between local and jev is a config change, not a rewrite claude does the hard reasoning. the local model picks allow, ask or deny. the hook enforces it. i break down the gate, the router and the real bill in the article below:
This post quotes another post: x.com/asteri_eth/status/2104901781725598103