Skip to content

Start Here

AegisAgent lets you run AI agents without losing control.

Overview

Before an AI agent does something risky, Aegis checks it, controls it, records it, and proves what happened. Aegis is not an agent — it is the control layer for agents.

Who it's for: teams putting AI agents near real credentials, real APIs, and real data — and the security people who have to answer for it.

Status: The known-agent integrity path is largely production-ready. Unknown-agent runtime enforcement and multi-replica enterprise operations remain Partial. Check Implementation Status before evaluating a specific claim.

Pick your path

🟢 "I am new"

  1. The_One_Minute_Tour.md — the whole idea in 60 seconds
  2. What_Is_AegisAgent.md — what it is and is not
  3. How_It_Works.md — the mechanism, step by step
  4. Then, when curious: Last_Mile_System_Walkthrough.md — the full story with code

🔵 "I am a developer"

  1. onboarding/For_SDK_Developer.md — protected tool calls in ~10 minutes
  2. Local_Development.md — run the whole stack locally
  3. flows/Known_Agent_Flow.md — what actually happens on every call
  4. Reference: components/SDK.md · api-reference.md

🔴 "I am security / SOC"

  1. onboarding/For_Security_Architect.md — trust boundaries in ~15 minutes
  2. onboarding/For_SOC_Analyst.md — investigate an incident in ~10 minutes
  3. AegisAgent_Threat_Model.md — the adversary's view
  4. Implementation_Status.md — the honest implemented-vs-planned ledger

Other roles — frontend, DevOps, new maintainer: onboarding/ has a guide each. The complete map of every document is Documentation_Index.md.

The three documents everyone should read

  1. Product_Overview.md — what it is, why it exists (10 min)
  2. Last_Mile_System_Walkthrough.md — the entire system as one story (15 min)
  3. Architecture_Overview.md — control points, trust boundaries, fail-closed rules (15 min)

Explore visually

One honest rule

Aegis controls what passes through Aegis control points (SDK, gateway, policy, approvals, receipts, SOC, MCP gateway — and, on the roadmap, tool broker, egress proxy, agent cage, node sensor). An agent that bypasses every control point is treated as unsafe: in runtime-controlled modes it is isolated, blocked, killed, quarantined, or banned. Nothing in these docs claims otherwise — check Implementation_Status.md before believing any capability.

Naming note

Some common titles map to canonical files to avoid duplicates: API reference → api-reference.md (generated) · deployment → deployment-guide.md + production-hardening.md · runbook → runbooks/index.md · contributor guide → CONTRIBUTING.md · FAQ → faq.md · glossary → Glossary.md.

Try the Proof

make doctor
make demo

The proof must block an untrusted merge, reject an approval hash swap and replay, and verify the receipt chain. Continue with Quickstart for expected output and troubleshooting.