Aegis

The firewall for AI agents.

Aegis is an open-source firewall for AI agents. It sits on the execution path between agents and tools, classifies tool calls, enforces policies before execution, supports human approval flows, and keeps a tamper-evident audit trail for later review.

I contributed to Aegis and feature it here because it is closely aligned with our work on agent safety, guardrails, AI auditing, and trustworthy deployment of agentic systems.

GitHub Repository | Star on GitHub | Demo | Back to Open Source

If this project is useful in your workflow, please star the GitHub repository to help more practitioners discover it.

Why this project

Agent systems make high-speed tool decisions without a human in the loop by default. That creates practical risks around unsafe commands, prompt injection, sensitive file access, unintended data exfiltration, and weak auditability. Aegis is designed as a runtime control layer for these deployment-time risks.

Core capabilities

  • Pre-execution policy checks for tool calls before they are executed.
  • Human-in-the-loop approval flows for higher-risk actions.
  • Tamper-evident audit trails for security, compliance, and post-incident review.
  • Support for dashboards, policy management, and operational monitoring around agent behavior.

Why it matters for agent safety

Aegis complements model-side safety work with runtime enforcement. It is relevant when teams need guardrails not only at the prompt layer, but also at the tool invocation and policy enforcement layers for real deployments.

Project highlights

  • Designed for policy enforcement across modern agent frameworks and tool integrations.
  • Includes approval workflows, policy evaluation, and operational visibility through a compliance cockpit.
  • Supports practical deployment paths, including Docker-based setup and multi-language SDK integration.

Demo

A quick visual walkthrough is available in the repository: Aegis demo GIF.