Show HN: Fence – Jiminy Cricket for AI coding agents
TL;DR
Fence is an open-source tool from hoop.dev designed to stop AI coding agents before they run potentially destructive shell commands. The target cases include rm variants and sudo delete commands aimed at a user’s home directory, especially when an agent is confused or badly prompted. Instead of acting as a simple denylist, Fence claims to infer the intent behind a command rather than matching only exact strings.
Nauti's Take
The direction is right: agents need brakes at the execution layer, not just better prompts and optimistic system messages. The launch is still very early and PR-heavy, with little external signal and a big claim around intent detection.
The real test is whether Fence can explain why it blocks a command and avoid slowing down legitimate developer workflows.
Briefingshow
AI coding agents increasingly get direct access to terminals, repositories, and local files. That shifts the risk from code quality alone to operational safety: one misunderstood command can damage a project or user directory. Fence targets the gap between careful prompting and actual system-level guardrails.
Sources