Docs

Documentation

Full documentation lives with the repository. Start here to get oriented.

Get started

Quickstart & Demo

Run the end-to-end demo: ALLOW, BLOCK and REVIEW workflows.

System Architecture

How the gateway, engines and audit trail fit together.

Core concepts

Policy Model

Rules, conditions, effects (ALLOW/REVIEW/BLOCK) and precedence.

Threat Model

What Aegis defends against — and what it deliberately does not.

API Contract

The REST surface: evaluate, registries, policies, approvals, audit.

Integrations

Python SDK

aegis_sdk: typed client with ActionBlockedError and ApprovalRequiredError.

MCP Security Gateway

Zero-trust wrapping for Model Context Protocol servers.

VS Code Extension

Approvals and monitoring from inside the IDE.

Attack Lab

Validate controls against adversarial scenarios.

Frequently asked

What is Aegis?

A runtime security gateway for AI agents. It sits between autonomous agents and their tools, evaluating every action before it executes through deterministic policies, risk scoring and threat detection — independent of the LLM.

How does Aegis protect AI agents?

By intercepting actions before execution and evaluating them through permission, trust, policy, risk and threat engines. The LLM is never the authority on authorization, evaluation fails closed, and every decision is audited.

Does Aegis work with browser-based agents?

Aegis protects any agent that routes its actions through the SDK, REST API or MCP gateway — including agents that browse the web. It does not ship a browser extension today; the shipped extension targets VS Code.

What happens when a threat is detected?

Depending on severity and policy: the action is allowed, paused for human approval (REVIEW), or blocked outright. Every outcome — including the reasons — is recorded in the audit trail, and REVIEW requests wait in the approval queue.

Can I create security policies?

Yes. Policies are rules with conditions and effects (ALLOW, REVIEW, BLOCK), evaluated with strict precedence via the API or the dashboard. The default is deny when nothing matches.

Does Aegis store prompts or browsing information?

Aegis records action metadata (agent, tool, operation, resource, environment) and redacted parameters. Sensitive values are replaced before logging. It does not record full prompts or browsing history.

Which environments are supported?

Any environment your agents run in — actions carry an environment field, and the risk engine applies stricter multipliers to production. Aegis itself runs via Docker (PostgreSQL + Redis) or locally with uvicorn.

Is Aegis open source?

Yes. The gateway, dashboard, SDK, MCP gateway and VS Code extension are all on GitHub at Bham0603/Aegis.