Before an agent gets your shell, review the instructions it was handed — AEGIS

I maintain AEGIS at Escape Studio; AI assisted development and this post. AEGIS is a free, MIT-licensed Windows-first security beta for local AI agents.

A skill is a set of instructions that can include shell commands and permission changes. The review question is practical: does this instruction ask the agent to skip a confirmation step, and where is that request in the file?

AEGIS can review a selected local skill, project or supported agent profile and show findings with their source context. The screenshot below illustrates an approval-bypass instruction in a simulated review fixture. That gives the operator a concrete passage to inspect before deciding whether to use the skill.

The next protection step is explicit configuration. A command routed through AEGIS’s execution policy can be denied before launch. On the configured Model Context Protocol gateway, a one-use approval matches the exact tool and arguments; changed arguments need a matching new grant. Process pause and stop are manual, with an identity check.

AEGIS Local security review, captured 27 October 2026 from the current source UI preview with simulated data. No live blocking test is depicted.

Captured 27 October 2026 from the current source UI preview; simulated data, no live blocking test.

Public beta: 0.20.1-beta. Windows is the primary platform; macOS/Linux are experimental. The enforcement controls cover configured routes.

Source, beta and contribution guide: GitHub - antropos17/Aegis: Windows-first desktop oversight for AI agents: live monitoring, local security reviews, token usage, audit reports and opt-in action policies/MCP gateways. Experimental Hyper-V guest execution; macOS/Linux experimental. · GitHub

When reviewing a skill, what source context would help you decide whether its command is appropriate for your project?