See what your AI agents can touch.
One command scans Claude Code, Cursor, Copilot, and every MCP server on your machine — then hands you a shareable report of every path they can read, write, or execute.
Join the early access list for the report generator. No spam, one email when the scanner is ready to run.
Why AI Agent Permissions Matter Right Now
A project called Geiger hit the front page of Hacker News this week with a simple pitch: show every AI agent running on your machine and what it can touch. It earned 44 points and a small but unusually focused comment thread — and the same conversation surfaced independently on three separate platforms. That pattern matters. When developers on unrelated forums start asking the same question in the same week, it usually means a real operational anxiety has crossed the threshold from curiosity into daily friction.
The reason is straightforward. Over the last year, most working developers quietly accumulated a stack of agents: Claude Code with a settings file full of allow rules, Cursor with its own permission model, a handful of MCP servers wired into project roots, and maybe a Copilot config nobody has opened since install. Each tool shows you its own permissions in its own panel. None of them show you the union. So the honest answer to "what can my agents actually reach on this laptop?" is that nobody knows — not the developer, and certainly not the security reviewer who has to sign off on it.
Existing solutions do not close that gap. Built-in commands like /permissions cover a single tool
and cannot be exported. Dependency scanners such as Snyk and Socket.dev audit packages, not agent file access.
Password managers guard credentials, not directory trees. What is missing