23 -- Harden a Public Authenticated MCP Gateway

Prerequisite: 22 -- External Multi-Tenant OIDC You will need: A working front-door deployment (recipe 22), plus platform ownership of your public edge (managed LB/WAF, DNS, certificates) Time: Read first, then plan -- this is a reference architecture, not a 20-minute run Adds: The trust boundary, control catalog, threat model, launch checklist, rollback plan, and incident-response handoff for exposing a Hangar front door to the public internet

The Problem

Recipe 22 proved the identity guarantees of a front door on your laptop: two tenants authenticate through OIDC and each sees only its own tools. That recipe is deliberately a local, test-only profile. Exposing the same gateway to untrusted traffic on the public internet needs a different class of guarantee -- a hardened network boundary, edge protection, durable and immutable audit, high availability, and a rehearsed incident path.

This recipe is a reference architecture and checklist, not a copy-paste deploy guide. Every config block below is illustrative and uses placeholders. The runtime controls Hangar owns are named against real config keys and environment variables; the edge controls it does not own are called out as your platform's responsibility.

Managed, publicly trusted TLS only. The self-signed certificates and local Keycloak in recipe 22 are an integration-test convenience only. A public gateway MUST terminate TLS with a managed or publicly trusted certificate at a load balancer, WAF, or reverse proxy you operate. Never expose a self-signed or locally minted certificate to the internet.

Security review gate. Any concrete public-edge configuration -- real domains, host allow-lists, CORS origins, header rules, certificates, secrets, and upstream rate limits -- MUST pass human security review before it is published or deployed. Nothing in this recipe is a substitute for that review.

The Trust Boundary

There are two networks and one boundary. Public traffic terminates at an edge you operate; Hangar and its providers run on a private network with no directly reachable service ports.

                    PUBLIC INTERNET (untrusted)
                              |
                    TLS (managed / publicly trusted cert)
                              |
        +---------------------v----------------------+
        |   EDGE  (external-infrastructure)          |
        |   managed LB / WAF / DDoS / edge rate cap  |
        |   terminates TLS, routes by host           |
        +---------------------+----------------------+
                              |  private network only
  ====== PRIVATE SERVICE NETWORK (no public ingress) ======
                              |
              +---------------v---------------+
              |   Hangar front door (HA)      |   tool_access.mode: front_door
              |   OIDC validation, RBAC,      |   auth.allow_anonymous: false
              |   per-tenant projection,      |   MCP_TRUSTED_HOSTS / CORS scoped
              |   rate limiting, audit        |
              +---+-------------------+-------+
                  |                   |
        per-provider service    durable stores
        accounts (least priv)   (auth + event/audit)
                  |                   |
        +---------v------+   +--------v---------+
        |  MCP providers |   |  auth store +    |
        |  (private)     |   |  event store     |
        +----------------+   +------------------+

The single load-bearing idea, carried over from the Deployment Playbook: the topology mode and auth posture are the boundary, not the network. The private network is defense in depth. The thing that actually holds the line is tool_access.mode: front_door with auth.allow_anonymous: false, which denies an unauthenticated caller before any backend is consulted (recipe 22, step 3).

Verify the boundary

The boundary is only real if you can prove it. Two checks:

1. No direct service port from the public side. From an external host (not on the private network), a direct hit on Hangar's service port must not connect. Only the edge's TLS port answers.

   # From outside the private network: the service port is unreachable.
   nc -z -w3 hangar.internal.example 8000 ; echo "exit=$?"   # non-zero = refused/filtered
   # Only the public edge answers, and only over TLS.
   curl -sS -o /dev/null -w "%{http_code}\n" https://gateway.example/health/live

2. Anonymous traffic is denied at the door, not proxied. Through the edge, an unauthenticated MCP call fails closed:

   curl -s -o /dev/null -w "%{http_code}\n" https://gateway.example/mcp \
     -H "Content-Type: application/json" \
     -d '{"jsonrpc":"2.0","method":"tools/list","params":{},"id":1}'   # -> 401

   gateway.example, hangar.internal.example, and the port are placeholders. Substitute your reviewed values.

Control Catalog (who owns each control)

Every control is classified by who is responsible for it:

• application -- a Hangar config key or environment variable you set.
• provider -- a property of a backend MCP server or the identity provider.
• external-infrastructure -- your platform: LB/WAF, DNS, managed database, SIEM, orchestrator.

Illustrative Config (placeholders -- review before use)

This shows only the shape of the production knobs on top of the recipe-22 front door. It is not deployable as written.

# config.yaml -- ILLUSTRATIVE. Every value is a placeholder pending security review.

tool_access:
  mode: front_door                 # untrusted edge: fail-closed, flat per-tenant tools

auth:
  enabled: true
  allow_anonymous: false           # deny unauthenticated callers before any backend
  oidc:
    enabled: true
    issuer: https://idp.example/realms/prod          # your production IdP
    jwks_uri: https://idp.example/realms/prod/protocol/openid-connect/certs  # HTTPS
    resource_uri: https://gateway.example            # advertised AND enforced as aud
    tenant_claim: tenant_id

  # Durable, backed-up auth storage. `postgresql` for HA (all replicas share it);
  # `sqlite` on a durable volume is acceptable for a single node.
  storage:
    driver: postgresql             # memory | sqlite | postgresql
    host: db.internal.example      # managed, private, backed up
    port: 5432
    database: mcp_hangar

# Durable audit trail -- fail fast rather than silently lose history.
event_store:
  enabled: true
  driver: sqlite
  path: /app/data/events.db        # writable, backed-up volume
  allow_memory_fallback: false

# In-process rate limit. This is NOT your DDoS defense -- the edge WAF is.
rate_limit:
  rps: 50                          # placeholder; size against real capacity
  burst: 100

mcp_servers:
  reports:
    mode: remote
    endpoint: https://reports.internal.example/mcp
    tls:
      verify: true                 # never disable backend TLS verification
    auth:
      type: bearer
      token: ${REPORTS_PROVIDER_TOKEN}   # per-provider service account, least privilege

    tool_access:                   # tenant-scoped, read-only surface
      member:
        "tenant:acme":
          allow_list: [list_reports, get_report]   # read-only tools only

The public-edge pieces -- the LB/WAF, the TLS certificate, host routing, edge rate caps, DDoS rules, and the SIEM pipeline -- are intentionally absent here. They are external-infrastructure, and their concrete configuration is exactly what security review must approve before publication.

Try It: Tenant-Scoped Read-Only Provider + Global Withdrawal

You can rehearse the two guarantees that matter most at the edge against your recipe-22 stack (local, test-only TLS). The commands are the same over the public edge; only the host and scheme change to your reviewed HTTPS endpoint.

1. Confirm a tenant sees only its read-only surface. With the reports provider's allow_list of list_reports, get_report for tenant:acme:

   curl -s https://gateway.example/mcp \
     -H "Authorization: Bearer $ACME_JWT" \
     -H "Content-Type: application/json" \
     -d '{"jsonrpc":"2.0","method":"tools/list","params":{},"id":1}' \
     | jq -S '.result.tools[].name'

   Expected -- a flat, read-only surface, no meta-API, no other tenant's tools:

   "get_report"
   "list_reports"

2. Prove a tool the tenant may not call is refused even by name (the projection hides it and the call path re-checks it):

   curl -s https://gateway.example/mcp \
     -H "Authorization: Bearer $ACME_JWT" \
     -H "Content-Type: application/json" \
     -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"delete_report","arguments":{}},"id":1}' \
     | jq '.error.code // .result.isError'   # -> -32601 (method not found)

3. Withdraw a tool globally at runtime and verify it disappears for every tenant. This needs an admin key with the lifecycle action on mcp_servers:

   curl -s -X POST \
     https://gateway.example/api/admin/tools/reports/get_report/withdraw \
     -H "X-API-Key: <admin-key>" \
     -H "Content-Type: application/json" \
     -d '{}' | jq

   Expected -- a global withdrawal (no tenant_id):

   {"withdrawn": true, "mcp_server": "reports", "tool": "get_report", "tenant_id": null}

   Re-run step 1: get_report is gone from acme's list, and a tools/call for it is refused at the enforcement path before any backend is contacted. restore with the same body reverses it. This is your emergency "pull a tool from the whole fleet" lever -- rehearse it before you need it.

Threat Model

State what you are defending against, and what you are not.

Out of scope / not claimed here: Hangar is an OAuth Resource Server; it validates tokens but never issues them -- IdP hardening is the provider's job. Hangar does not provide a mandatory independent policy gate. An external policy engine (OPA) can be wired in as an optional, future layer, but the runtime does not today enforce a composed "RBAC and OPA must both allow" decision, so do not present OPA as a required second gate. Physical/host security and the managed database's own hardening are external-infrastructure.

Launch Checklist

Walk this before opening the gateway to the public. The general promotion list lives in Cookbook 13 -- Production Checklist; this adds the public-edge items.

• TLS: managed/publicly trusted certificate at the edge; no self-signed cert exposed; backend mcp_servers.<id>.tls.verify: true.
• Network boundary: no public ingress to Hangar's service port; only the edge is reachable; the nc/curl boundary checks above pass.
• WAF / DDoS: edge WAF and DDoS protection enabled with an edge rate cap; Hangar rate_limit.rps / rate_limit.burst set as a backstop.
• Identity: tool_access.mode: front_door, auth.allow_anonymous: false, OIDC issuer/jwks_uri (HTTPS)/resource_uri set; anonymous call returns 401; wrong-audience token rejected; MCP_JWT_MAX_TOKEN_LIFETIME set.
• Host / CORS: MCP_TRUSTED_HOSTS and MCP_CORS_ORIGINS replaced with reviewed production values (never the dev defaults); MCP_TRUSTED_PROXIES set to your proxy chain.
• Storage: auth.storage on a managed, backed-up database; event_store durable with allow_memory_fallback: false; backups verified restorable.
• Logging / SIEM: MCP_JSON_LOGS=true, MCP_LOG_LEVEL=INFO; logs shipped to an append-only, access-controlled SIEM.
• HA: multiple replicas behind the LB sharing the same durable stores; orchestrator gates on /health/live and /health/ready.
• Per-provider service accounts: each provider has its own least-privilege credential; no shared super-credential; tenant surfaces are least-privilege.
• Incident readiness: global tool-withdrawal rehearsed; rollback rehearsed; on-call and SIEM alerts wired.
• Security review: the concrete public-edge configuration is approved.

Rollback Plan

A public rollout must be one reversible step, not a rebuild.

• Pin, don't chase. Deploy a pinned image digest (not latest) and keep the previous digest and the previous config.yaml one revert away.
• Roll back at the edge first. The fastest safe reaction to a bad rollout is to shift the LB back to the last-good replica set; the private network makes this a routing change, not a redeploy.
• Withdraw before you roll. If a single tool or provider is the problem, use the runtime global withdrawal (Try It, step 3) to pull it fleet-wide in seconds while you prepare the full rollback.
• Health-gate the return. Bring the previous version back behind /health/ready; hold traffic until readiness is green so a node that lost durable storage is never routed to.
• Config reload, not restart, for scoped changes. Tightening a tenant's tool_access or withdrawing a tool is a reload-driven overlay; you do not need to restart the fleet to narrow the blast radius.

Incident-Response Handoff

When something is wrong at the edge, the responder needs a known path:

1. Contain. Withdraw the affected tool(s) globally (Try It, step 3) and/or shift the edge away from the affected replicas. Rotate the OIDC signing keys at the IdP and any leaked per-provider credential; Hangar re-fetches JWKS on the new kid without a restart.
2. Preserve evidence. Freeze and snapshot the durable event store and the SIEM window covering the incident before any redeploy overwrites state.
3. Attribute. Use the structured logs and metrics (below) to bound which tenants, providers, and tools were involved.
4. Recover. Roll back per the plan above; restore the affected provider's least-privilege credential.
5. Review. Feed findings back into the threat model and launch checklist.

Name the on-call owner, the SIEM query location, and the credential-rotation runbook here for your deployment -- those are environment-specific and belong in your internal runbook, not in this reference.

Evidence Collection

Everything a responder or auditor needs is observable at the boundary:

• HTTP responses are first-line evidence: 401 for anonymous/wrong-audience calls, JSON-RPC -32601 for a projection miss, and a CallToolResult with isError: true for a call denied at the enforcement path. No denied call reaches a backend.
• Structured logs (JSON) carry the wiring and per-request surface: OIDC enablement, per-tenant policy loads, and per-request tenant_id and tool counts. Shipped immutably to the SIEM, they are the durable audit record.
• Metrics at /metrics quantify the boundary over time: mcp_hangar_tool_access_denied (authorization refusals), mcp_hangar_rate_limit_hits (throttling pressure), mcp_hangar_tool_calls_total (call volume by outcome), mcp_hangar_mcp_server_up and mcp_hangar_health_checks_total (provider and gateway health). Alert on denials and rate-limit spikes as early indicators of probing.
• The durable event store is the tamper-evident trail of state changes (tool withdrawals/restores, key events); with allow_memory_fallback: false it either persists or the node is pulled from readiness.

Key Config Reference

What's Next

This recipe is the public-edge capstone of the Deployment Playbook series (mcp-hangar/docs#17). To go deeper on the pieces it assembles:

• Identity and per-tenant isolation end to end -- 22 -- External Multi-Tenant OIDC.
• Multiple IdPs, one per customer -- 17 -- Multi-Issuer Front Door.
• Sandboxed, read-only providers with controlled writes -- 20 -- Read-Only Rootfs & Controlled Writes.
• The stage-by-stage maturity model -- Deployment Playbook.
• Pre-launch fundamentals -- 13 -- Production Checklist.

Before you publish any concrete public-edge configuration, route it through security review. This document is the architecture and the checklist; the approved, environment-specific edge config lives behind that review.