Open-source · MIT · v0.1.0
Every tool call — classified, gated, audited. Zero code change.
Same code either way. Watch the 90-second tour → · Compare the two paths
Works with the agents you already ship — incl. LangGraph, AutoGen, Mastra
What you see
One dashboard surfaces every agent decision, every block, every anomaly — across every workflow you ship.
What it does
Pre-deploy scan
Point AEGIS at any agent codebase. Tree-sitter AST across Python / JS / TS finds every tool call, every framework, every credential — and proposes a starting policy before the first deploy.
Plain English → Policy
No DSL to learn. Type one sentence — "block emails to personal addresses during checkout" — and AEGIS emits a grammar-constrained policy your gateway can enforce instantly. Auditable, reversible, version-controlled.
Describe the rule
Block emails to personal addresses during checkout flow. Allow ops@acme.io but flag anything to gmail, outlook, or icloud.
↓ ✨ Generate
rule: "block-personal-email-in-checkout" when: - tool.name == "send_email" - context.workflow == "checkout" recipient: deny: ["@gmail.com", "@outlook.com"] allow: ["@acme.io"] action: BLOCK
Runtime block
The same gateway that serves your agents enforces every policy. Allow, escalate, block — decided before the call ever leaves your network. PII redacted, anomalies surfaced, every decision cryptographically chained.
Forensic audit
Every block lands in a Merkle-chained log. Group by policy to find the rule that's actually firing. Filter by CRITICAL / HIGH / MEDIUM / LOW. Hand the same pack to your auditor that you'd hand to an incident responder.
Agent registry
Every agent registers once and stays accountable: status, owner, scope, secret rotation, last-seen environment. One place to suspend a misbehaving agent, rotate a key, or grant a new scope.
Coverage
Pull up a per-agent coverage report at any time: which tools are policy-gated, which categories have only audit coverage, which CVE detectors are subscribed. No mystery gaps when the auditor asks "what's enforced?"
5-minute integration
import openai
client = openai.OpenAI(
api_key="sk-xxxx"
)OPENAI_BASE_URL=https://gateway.aegis.dev/openai/v1
AEGIS_API_KEY=aeg_xxx
# code unchangedvs. the category
| Capability | AEGIS | Others |
|---|---|---|
| Cryptographic audit (Merkle + witness) | RFC 6962, built-in | none |
| Sequence-aware anomaly | n-gram LM, per-agent | single-call only |
| Multi-agent collusion | burst / relay / cycle | single-agent only |
| Workflow → per-node policy | 5 frameworks | none |
| Counterfactual explainer | verified by re-validation | partial |
| AST scan rules | tree-sitter + YAML | regex only |
| GenAI OTel semconv | full | proprietary |
| SCIM + SAML + OIDC | all three | one or the other |
| Policy effectiveness scoring | P/R/F1 + retire signal | none |
| License | MIT | closed |