Aegis Gateway 2.0

Governance

“Governance” means two things for this project, and this page covers both:

  1. How Aegis participates in Apache Incubator governance: applicable Apache Software Foundation and Incubator policy, supplemented by the repository procedures linked below.
  2. What governance Aegis provides to the teams that deploy it: the controls an operator gets over what AI agents can do with their systems.

Part 1: Apache Incubator project governance

Aegis is an Apache Incubator project. Applicable ASF and Incubator governance, contribution, conduct, licensing, release, security, and voting processes take precedence over repository procedures. The repository procedures in AGENTS.md and CONTRIBUTING.md support that governance but do not replace it. In particular:

Part 2: what governance Aegis gives an operator

Every enforcement control below is configuration, applies to every backend in the same way, and fails closed. Advisory hooks, decision listeners, metrics, and trace exporters are isolated observers: their failures cannot grant access or break an enforcement decision.

MCP governance guideline explains how an operator applies these controls as a complete runtime baseline. It is Aegis guidance, not an Apache Software Foundation policy, release procedure, compliance certification, or replacement for a project’s own governance.

Question an operator asks Control Where
Who is calling, on whose behalf? Host identity (JWT/JWKS, Knox Subject), effective subject, delegation chain, issuer ARCHITECTURE.md section 3
What may they even see? Per-group exposure of tools, resources and prompts; write tools hidden from readers access, allowedGroups
What does each tool do? Operator classification READ / MUTATE / DESTRUCTIVE; unclassified means DESTRUCTIVE; upstream hints only raise tools.<name>.class
On which data? Typed resources extracted from arguments by JSON pointer (spark:table:sales.orders), never widened to * resources, scopes
Is it allowed? Rules with deny-overrides and default deny, composable with OPA policy
Does the token carry the right grant? Per-tool OAuth scopes the verified own-host JWT must carry, checked before policy tools.<name>.actionScopes
Should a human agree first? Per-call approval bound to the exact action, single use, separation of duties approval, admin API
What would it do? Dry run (gateway, backend-native validate, or a preview tool) with a signed receipt required before a destructive run vrp, dryRun
Where may it reach? Egress allow list over every destination and DNS answer; metadata ranges blocked egress
How much? Per-caller and per-tool rates, weighted hourly budgets, per-backend circuit breakers rateLimits, breaker
Can the tool surface change under us? Pinned tool and prompt definitions; changes and poisoned text quarantined until accepted; optionally a signed register of which servers may start catalog, serverCatalog
Can we stop one server fast in an incident? Signed break-glass order that quarantines or revokes named servers for up to 24 hours, applied at restart serverCatalog.breakGlass
What reaches the model? Output bounds, secret redaction, content screening, binary refused unless allowed output, screening
What happened, provably? Intent journaled before execution; hash-chained, verifiable offline; W3C trace ids audit, --verify-audit
Is it working? Prometheus decision metrics by tool, source and outcome metrics

Governance state never lives in an MCP session, so the same controls hold for the stateful 2025 protocol and the stateless 2026-07-28 protocol, and for both front doors (the gateway’s own server and Knox).

What Aegis does not decide: whether an agent’s intent is good. It governs actions, data flow and accountability; it limits prompt injection but does not claim to prevent it (see SECURITY.md).

Which of these controls answer a regulatory question

GOVERNANCE-MAPPING.md takes the same controls and maps them to the EU AI Act, GDPR, DORA and NIS2, NIST AI RMF, SOC 2, ISO/IEC 42001 and 27001, the OWASP LLM and Agentic top tens, and a pointer table for other jurisdictions. Each control there names the code that implements it and a test that fails if it stops working. It is a control mapping to help an audit, not a compliance claim: Aegis has no certification, attestation or conformity assessment, and that page says so first. Read its gap table before repeating any coverage statement from it.