This document is the Aegis operator guideline for governing Model Context Protocol (MCP) interfaces. It describes the controls a project or deployment team should define before an agent can discover, invoke, or receive results from an MCP capability.
It is not an Apache Software Foundation policy, a release procedure, legal advice, or a compliance certification. Apache projects govern themselves through their PMCs and remain responsible for their own identity, privacy, security, release, and operational decisions. An Aegis deployment provides a technical control boundary and evidence, not a conclusion that a project complies with a law, standard, or Foundation requirement.
MCP governance is the set of controls that determines whether a caller may discover, invoke, delegate, approve, and audit an MCP capability. The decision must use the action’s semantics, not only the HTTP route or an authenticated client identity:
MCP governance complements, rather than replaces, transport security, API gateways, backend access control, model-provider controls, and project governance. Those layers must continue to perform their own roles.
Aegis applies this guideline at the MCP execution boundary. It does not decide whether an agent’s business goal or intent is good, prevent every prompt-injection attack, replace a least-privilege backend credential, or control data after a client receives it.
An Aegis deployment should apply these principles together. No single control is sufficient.
| Principle | Required operator outcome | Aegis control boundary |
|---|---|---|
| Semantic decision | A decision is made for a named MCP operation, its arguments, and typed resources. | Tool descriptors, schemas, typed-resource extraction, policy, and action binding. |
| Least privilege | Access starts denied and expands only through explicit policy and narrow backend credentials. | Default deny, deny-overrides, groups, scopes, resource rules, and source credentials. |
| Bounded delegation | An agent may act only with explicit, limited authority associated with an authenticated principal. | Host identity, effective subject, delegation chain, route groups, OAuth route scopes, and per-tool action scopes. |
| Human control | Consequential changes require independent human review before backend execution. | Per-call approval and destructive-operation preview receipts. |
| Trusted capability surface | A changed or unreviewed tool definition cannot silently change the available authority. | Catalog pinning, quarantine, source trust state, explicit acceptance, a signed server catalog, and time-bounded break-glass containment. |
| Safe data flow | Untrusted content cannot expand authority; sensitive content is bounded and protected. | Schema validation, resource validation, screening, output bounds, redaction, and egress policy. |
| Fail closed | An unavailable, ambiguous, or invalid control prevents execution. | Fixed governance order, hook failure denial, audit intent before execution, and strict configuration. |
| Verifiable accountability | The operator can reconstruct the decision, approval, and outcome for governed actions. | Intent-first hash-chained audit records and offline verification. |
Before exposing an MCP capability through Aegis, its owning team should document the following items. A missing item is a reason to delay exposure or classify the capability more strictly.
Each exposed tool, resource, prompt, or completion path should have:
READ, MUTATE, or DESTRUCTIVE;In Aegis, an unclassified proxied tool is DESTRUCTIVE. That conservative default prevents a
newly discovered capability from receiving write-like authority by omission. A source supplies
capability only. The operator supplies classification, resource taxonomy, approval rules, and
policy.
The deployment should authenticate the caller at the host boundary and preserve an accountable identity record containing the actor, effective subject, groups, issuer, and any delegation chain. Authentication establishes who presented a token; it does not itself authorize a tool call.
For delegated requests, the deployment should define the delegation issuer, audience, scope, expiry, and revocation behavior. A token valid for one MCP server or action class should not silently become authority for another. Aegis own-host deployments can bind OAuth resource metadata, audiences, route groups, and scopes at the perimeter, then apply tool and resource policy inside the governance engine.
Do not forward the caller’s credential to a backend as a substitute for semantic authorization. Aegis resolves the backend credential only after it allows and journals the action. The backend credential must still be least privilege for the source it serves.
Authorization should evaluate the complete governed operation:
kafka:topic:orders or iceberg:table:sales.orders.A policy that allows POST /mcp or a broad backend service account is not a substitute for this
decision. Aegis starts with default deny, gives explicit denies precedence, and refuses missing,
malformed, wildcard, or widened typed resources rather than converting them to broad access.
The operator should identify state-changing and high-impact operations before exposing them. Aegis provides the following baseline:
| Action class | Minimum Aegis control |
|---|---|
READ |
Standard identity, exposure, schema, typed-resource, policy, egress, screening, and audit decision path. |
MUTATE |
The READ controls plus independent, single-use, per-call approval bound to the exact action. |
DESTRUCTIVE |
The MUTATE controls plus a fresh preview or dry-run receipt bound to the same action. |
Approval and receipt binding must include the caller, tool, source, class, typed resources, argument digest, and tool-definition digest. A requester and its delegated identity must not approve their own action. A preview is not a valid receipt if it cannot safely represent the backend action that will follow.
A project may define a narrower exception for low-risk automation, but it should be public, reviewed, reversible, and backed by clear evidence. Aegis does not make autonomous state changes safe merely because a client or tool labels them routine.
MCP clients often discover tool names, descriptions, schemas, resources, and prompts dynamically. Those definitions are security-relevant because they influence both agent behavior and the available action surface. The deployment should:
A pin detects changes after an initial definition is trusted. It does not prove that a first observation was safe. Initial source review remains an operator responsibility.
MCP arguments, tool results, resources, prompts, tickets, logs, and retrieved documents may all contain attacker-controlled instructions. Treat this content as data, never as authority to alter policy, credentials, approval requirements, source trust, or egress rules.
An Aegis deployment should:
A remote hook is a separate data recipient. By default, Aegis sends it identity, operation, typed resources, and an argument digest, not raw arguments or results. An operator must explicitly opt in before such content leaves the gateway.
For every governed call, record enough information to explain the decision without indiscriminately recording sensitive raw content. The evidence should identify the caller, effective subject, delegation context, tool, source, action class, typed resources, argument digest, policy result, approval/receipt references, tool-definition digest, outcome, timestamp, and trace identifier.
Aegis writes an ALLOW_INTENT event before resolving the backend credential or calling the
backend. If that audit intent cannot be recorded, the call does not execute. Its local hash chain
and offline verifier support detection of local-history tampering, but do not replace retention,
access control, monitoring, alerting, or external audit anchoring. Those remain deployment
responsibilities.
The operator should define an incident process for revoked identities, compromised backend credentials, suspect tool definitions, audit-verification failures, denied-call investigation, and MCP-specific vulnerability disclosure. Apache software vulnerabilities continue to use the Foundation’s coordinated security process where applicable.
Use this workflow for a new source or an existing MCP server.
DESTRUCTIVE; relax
only after evidence supports a lower class.serverCatalog), establish catalog pins, and
define who can accept an intentional definition change.The illustrative Apache-project governance profile shows a configuration shape for a project community to adapt. It is not an ASF policy, release checklist, or proof of compliance. The configuration guide and adoption guide explain the configuration and validation boundaries.
| Responsibility | Aegis can enforce or evidence | The deployment or project must own |
|---|---|---|
| Perimeter | Host identity, route groups/scopes, request bounds, CORS configuration. | IdP lifecycle, credential issuance, TLS termination, network topology, and recovery. |
| Action governance | Tool class, typed resources, policy, approval, receipt binding, catalog integrity, and decision audit. | Accurate classification, resource taxonomy, approver roles, policy review, and source admission. |
| Backend safety | Blocks a request before execution and resolves credentials only after governance succeeds. | Least-privilege backend roles, native authorization, transaction safety, sandboxing, and rollback. |
| Data protection | Schema limits, egress enforcement, screening, content bounds, and recognized-secret redaction. | Data classification, retention, privacy review, key management, DLP, model-provider terms, and downstream handling. |
| Evidence | Intent-first local hash chain, structured decision records, and offline verification. | Secure storage, retention, off-host anchoring, access review, alerting, incident response, and legal obligations. |
ASF policy already establishes the project-level controls that matter to an Apache community: public and archived technical decisions, independent PMC governance, secure vulnerability handling, release voting, source and signature verification, licensing, privacy, and contributor accountability. The official ASF policy catalog does not establish a Foundation-wide runtime MCP-governance profile.
This guideline fills only the runtime gap for an Aegis deployment. It does not change ASF policy, replace a PMC’s technical judgment, or require any Apache project to adopt Aegis. A project that chooses Aegis should govern its configuration and operational exceptions through its normal public processes, while keeping embargoed vulnerability information private under the applicable security procedure.
Relevant ASF sources:
The source material is summarized and rephrased for compliance with licensing restrictions.
An Aegis deployment can demonstrate that configured controls load, deny or allow defined cases, record decisions, and work against a retained target version. That evidence is valuable, but it is not certification, legal advice, a security guarantee, or proof of compliance with a regulation or an ASF requirement.
Use VALIDATION.md, SECURITY.md, THREAT-MODEL.md, evidence/SECURITY-VALIDATION.md, GOVERNANCE-MAPPING.md, and the deployment-specific threat model to state exactly what was tested, what controls are configured, and which responsibilities remain outside the gateway.