Fence – Jiminy Cricket for AI coding agents

Name: Fence – Jiminy Cricket for AI coding agents
Author: andriosr

by andriosr·Jul 7, 2026·1 point·0 comments

View on HN

AI Analysis

●●SolidShip ItBig Brain

Intention-reading guardrail for AI agents, not just a command string denylist.

Strengths

•Reads command intention semantically instead of relying on brittle string patterns.
•Already deployed in production with Claude Code and Codex at hoop.dev.
•Open source on GitHub with active team seeking adoption feedback.

Weaknesses

•Currently supports Claude Code and Codex, Cursor integration still pending.
•Agent safety category is emerging with limited benchmarks for effectiveness.

Post Description

Hi everyone, I'm Andrios, founder of hoop.dev (YC W21).

2 weeks ago, I told our eng team they could spend 20% of their week on side projects. The result: They came out with 3 open source projects. Tools we're using ourselves.

Fence is the first one. It prevents an agent, confused or poorly prompted, to execute a catastrophic command before it happens. This could be 'rm -rf~’, ‘rm -fr~’ or even ‘sudo rm -rf $HOME’. It’s different from a denylist, because it reads the intention, not only the string.

We use this for Claude Code and Codex. We want to see adoption before jumping to Cursor next.

The team is very excited to release this, so please to drop any comments with feedback or issues you find:

github.com/hoophq/fence