Comparison
API Gateway vs ARIA Shield vs MCP Gateway
Three layers, three jobs. API gateways route traffic. ARIA Shield proxies AI model calls and enforces budgets. MCP Gateway authorizes what agents do at the tool-execution layer. Most production deployments run all three.
When an agent calls a tool, your API gateway sees an HTTP request. It routes and rate-limits. But it can't answer: Does this user's session have budget left? Is this agent authorized to call this tool with these parameters? Does this action match an approved plan?
ARIA addresses this with two purpose-built gateways — both backed by the same AuthZEN PDP for authorization decisions: ARIA Shield (the BFF that proxies all AI model traffic, enforces budgets and content security, and keeps tokens off the client) and MCP Gateway (the PEP that authorizes agent tool execution with schema pinning, plan discipline, and signed receipts). Both sit behind your API gateway.
What each layer does
Three complementary layers, not competing products
API Gateway
Network & traffic layer
Infrastructure concerns for all HTTP traffic — human, API, or agent.
ARIA Shield
BFF for AI models & browser sessions
The single authorized surface for all AI model traffic. When the PDP authorizes a call, Shield unvaults per-service tokens from the OAuth Vault, makes the downstream call, and returns data — tokens never reach the client. AuthZEN PDP + OAuth Vault
MCP Gateway
Tool-execution authorization layer
Understands what an agent is trying to do — the tool, the parameters, the plan — and authorizes before execution. AuthZEN PDP
API Gateway is your existing infrastructure (Kong, NGINX, AWS ALB, etc.). ARIA Shield and MCP Gateway are EmpowerNow components — both delegate to the same OpenID AuthZEN 1.0 PDP for authorization decisions, so one policy governs UI, API, and agent surfaces.
Side-by-side comparison
| Capability | API Gateway | ARIA Shield (BFF) | MCP Gateway |
|---|---|---|---|
| Routing & rate limiting | Core function | Delegates to API GW | Delegates to API GW |
| TLS termination | Core function | Delegates to API GW | Delegates to API GW |
| LLM proxy | No — model-unaware | Yes — single auth + billing surface for all AI model traffic | No — tool-call layer |
| Zero-token SPA | No | Yes — HTTP-only cookies, no browser tokens | Not applicable (agent workloads) |
| Per-route PDP decisions | No | Yes — AuthZEN decision per route | Yes — AuthZEN decision per tool call |
| Budget & streaming caps | No | Yes — HTTP 402, spend/stream limits | Yes — budget stops pre-execution |
| Content security | No | Yes — PDP obligations enforce classification and egress controls | Yes — parameter allowlists, anti-injection checks |
| AuthZEN PDP | No | Yes — same PDP as MCP Gateway | Yes — same PDP as ARIA Shield |
| Schema pin validation | No | No | Yes — hash/version tamper detection |
| Plan discipline | No | No | Yes — deny off-plan actions pre-execution |
| Token brokering | No | Yes — per-service tokens minted server-side | Uses ARIA Passport |
| Signed receipts | Access logs only | Yes — settled to actuals | Yes — hash-chained proof of every action |
Key concepts
What makes MCP Gateway different from network-layer enforcement
Schema pinning
Every tool the agent can call has a registered schema with a SHA-256 hash. Before execution, MCP Gateway verifies the schema hasn't been modified — preventing tool-tampering attacks where an adversary alters a tool's definition after approval.
Plan discipline
Agents propose a plan of actions. MCP Gateway enforces that each execution matches the approved plan — no off-plan tool calls, no parameter drift. If an agent tries to execute a step that wasn't approved, execution is denied before it starts.
AuthZEN constraints
Authorization evaluated per the OpenID AuthZEN 1.0 standard — fine-grained decisions about which agent, calling which tool, with which parameters, under which delegation context. Not just "is this user allowed" but "is this action authorized right now."
Signed receipts
Every action that passes through MCP Gateway produces a signed, hash-chained receipt — cryptographic evidence of what was authorized, what parameters were used, and what the result was. Not a log entry. Proof.
How all three layers work together
Two paths through the stack — both governed
Browser / AI model path
Agent / tool-call path
Frequently asked questions
Can I use both?
Yes — that's the recommended deployment. Your API gateway handles routing, rate limiting, and TLS. MCP Gateway sits behind it and handles authorization, plan enforcement, and receipts. They operate at different layers of the stack.
Does MCP Gateway replace my API gateway?
No. MCP Gateway doesn't handle routing, rate limiting, TLS, or load balancing. It adds a governance layer that understands agent tool semantics — something API gateways aren't designed to do.
What's the difference between ARIA Shield and MCP Gateway?
ARIA Shield is the BFF for AI model traffic and browser sessions — it proxies all LLM calls through a single authorized surface, enforces budgets and streaming caps, applies content security obligations (classification, egress controls), and keeps tokens off the client. MCP Gateway governs agent tool execution — schema pinning, plan discipline, and parameter-level authorization. Both delegate to the same OpenID AuthZEN 1.0 PDP for authorization decisions, meaning one policy governs UI, API, and agent surfaces. Both emit signed receipts.
Do I need MCP support to benefit?
MCP-native agents get the fullest benefit, but schema pinning and AuthZEN authorization decisions add value for any tool-calling agent — regardless of protocol.
What happens if the agent tries an off-plan action?
MCP Gateway denies execution before any side effects occur. The denial is logged with a receipt explaining what was attempted and why it was rejected — useful for debugging and audit.
How is this different from a WAF?
A web application firewall inspects HTTP payloads for known attack patterns. MCP Gateway understands agent intent — it evaluates whether a specific tool call with specific parameters is authorized under current policy. Semantic authorization, not pattern matching.
Go deeper
ARIA Shield →
BFF gateway: zero-token SPAs, budget enforcement, streaming caps.
MCP Gateway →
Tool-execution PEP: schema pins, plan discipline, constraint enforcement.
Runtime Execution Control →
The full ARIA governance pipeline — SSC, WAITING protocol, proof chain.
Zero-Token SPAs →
The BFF pattern that keeps tokens out of the browser.
See MCP Gateway in action
Watch an agent execute a governed tool call — from plan approval to signed receipt.