AgentPermit – guard AI agents' local file access

Stops local AI coding agents from touching keys and env files they shouldn't

Problem

More people let Claude Code and Cursor run commands and read files locally, but these agents can casually read SSH private keys and .env files. A malicious prompt or a wrong move can leak them. The community frames the risk as the lethal trifecta—data access, communication, exposure—yet ordinary users lack a local valve that gates access by file type and path.

Solution

A small Mac-native policy tool where users declare which files, by type, path, or Finder tag, agents may not read. When an agent process tries to access them, the tool blocks it and asks for explicit approval, while keeping an auditable access log. No full antivirus—just need-to-allow access.

Users

Indie developers and security-conscious engineers on Mac running AI coding tools against real projects

MVP

Configure protected paths and file types (SSH, env); block reads via macOS file monitoring; prompt for human approval on block; per-agent access logs; exportable policy templates

Revenue

Open-source core with paid policy template packs or team policy sync; or a one-time desktop license

Why now

Local AI agent execution is going mainstream while sensitive data sits on the same disk. Enterprise security suites dominate, leaving individuals without a lightweight, explainable local gatekeeper.