Kong Agent Gateway vs Portkey Agent Gateway

Reviewed 17 August 2026 · 15 minute read

Choose Kong when the enterprise wants MCP and A2A controls inside an established API-gateway estate with declarative configuration, enterprise plugins and existing traffic operations. Choose Portkey when the team wants model routing, agent traces and managed MCP access in one AI-focused control surface. Test the exact edition and feature status because the two products overlap but do not expose identical controls.

Neither gateway automatically authorizes the business meaning of a refund, payment or production change. Both can enforce an external transaction decision.

Kong Agent Gateway vs Portkey Agent Gateway by responsibility

Swipe horizontally to inspect the full diagram.

Kong's current documentation places its MCP and A2A features in Kong Gateway 3.14. The AI MCP Proxy plugin documents per-tool ACL behavior, while the A2A proxy documentation describes protocol parsing, card rewriting and observability. Several capabilities require Kong AI Gateway Enterprise or Konnect, so confirm topology and licensing.

Portkey's MCP Gateway overview describes centralized authentication, access control, credential injection and request logging. Its Agent Gateway announcement labels that broader product beta and describes registry, policy, routing and traces. The separate authorization page currently lists workspace and server control as available and tool-level control as coming soon. Treat that contradiction as a procurement question, not a wording detail.

Enterprise agent market positioning

Product-scope analysis, not market share, quality or maturity

Text summary of highlighted companies

Company or protocol	Primary scope represented on the map
Intelliger	Transaction-specific authority, deterministic enforcement, reconciliation and portable evidence.
Kong	LLM, MCP and A2A traffic gateway, authentication and observability.
Portkey	AI and MCP gateway, credential isolation, observability and guardrails.
Cloudflare	Distributed traffic, security, remote MCP and AI gateway infrastructure.
MuleSoft	Enterprise API integration, orchestration and agent-facing connectivity.

Kong is stronger when gateway operations already exist

Kong lets a platform team extend familiar gateway practices to agent traffic. Standard authentication plugins can protect an A2A route. MCP-specific plugins parse calls, expose metrics and enforce ACLs. Declarative configuration and existing rollout procedures can matter more than an AI-specific dashboard when hundreds of APIs already use Kong.

The tradeoff is product and topology complexity. A feature may exist only in a particular version, enterprise package, Konnect deployment or plugin scope. A proof of concept must match the production topology rather than a convenient hosted tutorial.

Portkey is stronger when model and tool operations belong together

Portkey combines model routing, fallbacks, budgets, guardrails and AI traces with MCP access. Its authentication documentation separates the agent-to-Portkey identity from the Portkey-to-server credential. That is useful when developers should use tools without receiving GitHub, Slack or internal API credentials.

Portkey's observability can correlate model and tool calls. That is operationally useful, but logging full parameters and responses can also capture secrets, personal data or regulated records. Test redaction, regional storage, retention and export before sending production payloads through the gateway.

Use one proof-of-concept fixture for both gateways

Deploy two MCP tools behind each candidate:

inventory.lookup(sku, location)        # read only
supplier.refund(orderId, amountMinor)  # consequential write

Create three identities: an inventory reader, a refund operator and an untrusted external agent. Configure the same IdP, server credentials and telemetry destination.

Run these tests:

1. Discover and call the lookup tool with the reader.
2. Attempt the refund tool with the reader and require denial before the provider call.
3. Change amountMinor after an external approval and require digest mismatch.
4. Send a token minted for another resource and require audience failure.
5. Attempt header spoofing for tenant and user identity.
6. Replay the same proof while retrying the same business operation with a stable idempotency key.
7. Stop the external policy service and verify the refund fails closed.
8. Return a large or malicious tool response and inspect size, content and prompt-injection controls.
9. Rotate the upstream credential without changing agent configuration.
10. Export logs and prove which fields were redacted.

Capture configuration, product version, license, provider-call count and the exact denial point. This turns a feature comparison into a repeatable evaluation.

Bind an external transaction decision to the request

The gateway can normalize a material tool call into an authorization request:

type TransactionAuthorizationRequest = {
  tenantId: string;
  subject: string;
  protocol: 'mcp' | 'a2a' | 'http';
  action: string;
  resource: string;
  protectedArguments: Record<string, unknown>;
  mandateRef: string;
  requestDigest: string;
  idempotencyKey: string;
};

type TransactionAuthorizationResponse = {
  decision: 'allow' | 'deny' | 'approval_required';
  requestDigest: string;
  expiresAt: string;
  obligations: Array<{ type: string; value: string }>;
};

Kong or Portkey authenticates the caller, strips untrusted identity headers and constructs protected fields from verified state. The external policy service evaluates mandate, amount, destination, approval and cumulative budget. An allow response expires quickly and binds to the request digest. The gateway then releases the upstream credential and records the provider reference.

Test outage and retry semantics

Where Intelliger fits

Intelliger should not recreate Kong's API operations or Portkey's model-routing plane. OATI can supply portable identity references, mandates, canonical transaction envelopes, deterministic decisions and receipts. A gateway plugin or external-authorization call enforces the decision at the traffic boundary.

That integration is most useful when the same business authority must survive an MCP call, an ordinary API, a payment provider and a later audit package. In this design, the gateway manages traffic while Intelliger preserves the exact transaction chain and its bounded evidence.

Questions platform teams ask

Is Kong better than Portkey for MCP?

Kong is often the better fit for teams extending an existing enterprise API gateway. Portkey is often the better fit for AI teams that want model and MCP operations together. The required topology, tool-level control and observability policy decide the result.

Does Portkey support tool-level authorization?

Its current official pages are inconsistent. The product overview describes tool provisioning, while the authorization reference says tool-level authorization is coming soon. Require a live test and written confirmation for the edition you plan to buy.

Can Kong govern A2A traffic?

Kong Gateway 3.14 documents an AI A2A Proxy for protocol parsing, card rewriting, logs, metrics and use with standard gateway controls. Check the required AI Gateway Enterprise license and supported deployment topology.

Can either gateway replace business approval?

Only if it evaluates and binds the exact business fields. Authentication, tool ACLs and rate limits are insufficient when approval depends on amount, supplier, destination, cumulative budget or prior outcome.

Current Intelliger and OATI boundary

OATI currently provides developer-preview schemas, signing, verification, deterministic mandate evaluation, replay controls, receipts, conformance fixtures and a reference Envoy path. The hosted trust and lookup slice is deployed.

The commercial Kong and Portkey plugins, managed gateway fleet, complete policy compiler, durable evidence worker, independent protocol review and production acceptance are incomplete. The integration contract is a target commercial capability over implemented OATI primitives.

Use the MCP gateway buying guide for the four-vendor evaluation, compare an MCP gateway with an API gateway and place the decision inside the enterprise agent control stack.

Expert review required before publication: a gateway specialist should rerun every test against the exact licensed versions and resolve the documented Portkey authorization-status conflict.

To prototype a deterministic decision behind either gateway, open the OATI developer path.