Have a small local model vet agent shell commands

Goal: Catch dangerous commands from my coding agent locally, without cloud calls.
Setup: agents used in source: Claude, Cursor; tools/MCPs stated: laya-guard.
Steps:
1. Run a small local server that classifies each tool call as allow, review or block; the agent’s hook calls it before execution.
2. Test on everyday commands: the first version blocked ls -la and npm run build, so add a careful allowlist.
3. Don’t allowlist by command name alone (git log --output can write files); split command chains (&&, |) and check every part.
4. Know the trade-off: it fails open if the server is down.
Rules: ask me before anything irreversible (sending, posting, paying, deleting); don’t invent data, results or tools I haven’t given you; tell me what you couldn’t verify.
Based on: Changsu Seong — https://dev.to/scs0209/i-let-a-322m-local-model-judge-every-shell-command-my-coding-agent-runs-2hf6

Steps are from the source. Includes a “Based on” credit line.