# AEGIS

A pre-execution firewall for AI agents that intercepts every tool call, classifies it against policy in real time, blocks violations and writes a tamper-evident audit trail, with one line of integration and no agent code changes.

AEGIS sits between an agent and its tools, so a call such as DROP TABLE users triggered by an ambiguous prompt, a hallucinated rm -rf /, or an unintended data exfiltration can be classified and blocked before it executes rather than caught afterward. A "Compliance Cockpit" gives operators six views: a rolling 24-hour activity overview with action, agent, blocked and pending-review counts; a forensic activity log per tool call with the exact policy and condition that fired; plus traces, policies, cost tracking, sessions and approvals.

It installs as a single downloaded binary for macOS, Windows or Linux, or a one-line curl script that verifies and places the binary automatically, with human-in-the-loop approval prompts for pending actions. The project positions itself against agent frameworks that let a model choose which tools to call, with what arguments, at machine speed, with no human in the loop and no undo.

Website: https://github.com/Justin0504/Aegis