Hi everyone, I'm Faiz, and I build Paveo. It started with a hook. I wrote a Claude Code hook to stop rm -rf, then found it would let the command through if the hook itself crashed. So I looked at public repos: searching GitHub for hooks that read certain environment variables, 86 of 153 settings files had at least one hook that can't work as written, mostly because it reads a variable Claude Code doesn't set. ZenML and an AWS samples repo have since merged fixes. Paveo is the checkpoint I wanted instead. Shell commands, file edits and tool calls are checked against your policy before they run. A refused call never happens, the agent is told why, and every decision lands in a tamper-evident log. It runs on your machine and sends nothing anywhere. Three ways in: - uvx paveo doctor checks the Claude Code hooks you already have, with nothing to set up - paveo init claude-code (or codex, cursor) installs the guard - paveo replay claude-code shows what it would have refused in your past sessions, before you turn it on. On mine: 173 rm -rf across 162 sessions. What it isn't: a sandbox. It reads commands as text, so a disguised command can get past it, and the README says exactly where. Where it's going: rules can't capture every judgment call. With our design partners we're building a model trained on your team's own decisions that flags what the rules allow but shouldn't, and someone on your team to approve what gets flagged. It can only make a decision stricter, never looser, so the proof stays intact. We're onboarding a small group now. If that's you, reply here or write to faiz@trypaveo.com. One question I'd love answers to: what would your security team need to see before letting an agent act in production?

Paveo

Stop risky AI agent actions before they run

Stop risky AI agent actions before they run