Brushed metal detail — industrial, precise — Shubham Dhage / Unsplash

Documentation

Five minutes to your first protected call.

A full docs portal at docs.aegis.dev launches with v1.0 (Mintlify-hosted). Until then, this page is the curated index — pick an install path, learn six concepts, jump to the deep references on GitHub.

Install paths

Four ways to integrate. They all converge on the same gateway. Pick the one that matches your stack and skip the others.

Hosted SaaS (zero code change)

Point your existing OpenAI / Anthropic / Mistral / Gemini SDK at the gateway URL. No code edits.

export OPENAI_BASE_URL=https://gateway.aegis.dev/openai/v1
export AEGIS_API_KEY=aeg_xxx
# upstream OpenAI key forwarded unchanged
# your code stays untouched

Desktop installer (macOS / Windows)

One-file installer. Embeds the gateway + cockpit + Node runtime. Local-only by default.

# macOS Apple Silicon (M1+)
open https://github.com/Justin0504/Aegis/releases/latest/download/AEGIS_0.1.0_aarch64.dmg

# Windows .msi build is unblocked — see /download for the latest.

Download page →

Docker Compose (self-hosted)

Full stack — gateway + cockpit + Postgres + Redis — on a single host or in your VPC.

git clone https://github.com/Justin0504/Aegis
cd Aegis
docker compose -f docker/compose.yml up -d
# → gateway on :8080, cockpit on :13003

Self-host guide →

SDK (Python / JS / TS)

In-process auto-instrumentation across 9 agent frameworks. One init() call.

# Python
pip install agentguard
# in your agent entrypoint:
import agentguard; agentguard.init(api_key="aeg_xxx")

# Node / TypeScript
npm install @agentguard/sdk
import { init } from '@agentguard/sdk'; init({ apiKey: process.env.AEGIS_API_KEY });

SDK reference →

Six concepts that explain everything else

AEGIS has six core ideas. Internalise these and the rest of the documentation is just "more buttons, same shape".

How AEGIS sees a tool call

Every call hits the classifier, runs through L1 rules + L2 ML + (optionally) L3 LLM-judge, matches the per-tenant Policy DSL, and ends with one of three decisions: allow, pending (human review), block. See /docs/policy-templates for how the DSL composes, /docs/api for the decision endpoint shape.

Policies vs DSL rules

Policies are JSON Schema validators that gate the call argument shape (e.g. SQL injection patterns, URL allowlists). DSL rules are the tenant-side overlay — "for THIS agent, ALSO require an approval if the call is HIGH risk during evening hours". Both run on every call; either can deny.

Agent identity (Agent ID v1)

AEGIS registers each agent and mints a signed JWT identity card. The SDK forwards X-AEGIS-Agent-Token; the gateway verifies. Lets policy + audit map every call to a stable agent identity even when an agent is renamed or migrated.

Audit & transparency log

Every decision is appended to a per-tenant RFC 6962 Merkle log. Tree heads are Ed25519-signed. Customers cache signed roots and verify inclusion offline via tools/verify-log/index.mjs. See Security → Cryptographic audit primitives.

Observability — GenAI OTel

Set OTEL_EXPORTER_OTLP_ENDPOINT to a Datadog / Honeycomb / Tempo / Grafana Cloud endpoint and every span ships with the full gen_ai.* attribute set + a gen_ai.guardrail.* extension for decision metadata. Zero-config dashboards on every major OTLP consumer.

Integration modes (pick one)

1) LLM Egress Proxy — zero code change, just env vars; 2) SDK auto-instrument — one init() call; 3) MCP proxy — change one config field for Claude Desktop / Cursor / MCP hosts; 4) HTTP gateway direct — for custom agents. Mix freely — they all converge on the same gateway.