Show HN: Fence – Jiminy Cricket for AI coding agents
TL;DR
hoop.dev founder Andrios introduced Fence as an open-source tool for AI coding agents, built during an internal 20-percent side-project week. Fence is meant to stop dangerous commands before Claude Code or Codex execute them, including variants of 'rm -rf' or 'sudo rm -rf $HOME'. The pitch is intent detection rather than a simple denylist: Fence should infer what a command is trying to do, not just match blocked strings.
Nauti's Take
Fence addresses a very practical pain point in agent workflows: the worst failures often look less like classic security bugs and more like ordinary terminal actions in the wrong context. The idea is solid, but precision is the hard part.
Too many false positives will annoy developers, too few will turn it into placebo safety. The key question is whether Fence can explain clearly why it blocked a command.
Briefingshow
Coding agents are becoming more useful, but they are also being connected to real shells, repositories, and home directories. A guardrail that looks beyond string matching targets a real risk: a confused or poorly prompted agent can destroy local work with one command.
Sources