Skip to content
Talk to Us

Blocks.ai publishes this comparison and sells in this market. Every claim about another product links to a public source.

DIY vs Kong AI Gateway: an in-house build or a gateway product

Build it yourself (DIY)

Building agent connections and controls in-house from open protocols, existing infrastructure, and open-source tools

Kong AI Gateway

Gateway that governs LLM, MCP, and agent-to-agent traffic

Short answer

DIY is not a product: you connect and govern agents yourself, with open protocols such as MCP and A2A that leave authorization logic to the implementer.⁠Source 1, Source 2 Kong AI Gateway is a product: one gateway for LLM, MCP, and A2A traffic; version 2.x pairs a Kong-managed control plane with data plane nodes you run.⁠Source 3, Source 4

By Blocks.ai teamLast checked

Where each one sits

Six layers of running AI agents at a company, and what each product’s own public sources say it covers.

These aren’t the same kind of product

Where they overlap
One DIY form stretches an API gateway to cover agents, and A2A’s docs highly recommend API management for externally exposed A2A servers.⁠Source 5 Kong says AI Gateway governs A2A traffic.⁠Source 3
Where they differ
DIY also covers what sits beside a gateway: service meshes such as Istio, private links such as AWS PrivateLink, and agent frameworks.⁠Source 6, Source 7, Source 8
Running both
Kong’s docs say AI Gateway proxies to an agent registered as an upstream URL.⁠Source 9 In the two write-ups cited, Uber built its own gateway and Pinterest checks tokens with Envoy.⁠Source 10, Source 11
Public sources · checked 2 October 2026
  • Offered
  • Preview
  • You build it
  • Not publicly documented

DIY

  • Build agents: You build itOpen-source frameworks like ADK⁠Source 8
  • Host and run agents: You build itSelf-hosted, like LangGraph⁠Source 12
  • Identity and access: You build itYour own identity provider⁠Source 13
  • Registry and governance: You build itYour own agent registry⁠Source 14
  • Traffic between agents, tools, and models: You build itYour gateway and service mesh⁠Source 10
  • Agents across organizations: You build itA2A with API management⁠Source 5

Kong AI Gateway

  • Build agents: Not publicly documented
  • Host and run agents: Not publicly documented
  • Identity and access: OfferedPer-agent allow or deny lists⁠Source 9
  • Registry and governance: PreviewKonnect Catalog agents⁠Source 15
  • Traffic between agents, tools, and models: OfferedGateway for LLM, MCP, A2A⁠Source 3
  • Agents across organizations: Not publicly documented

At a glance

TopicDIYKong AI Gateway
What it isAn in-house build on open protocols such as A2A and MCP, which the A2A specification calls complementary.⁠Source 1A gateway that governs LLM, MCP, and A2A traffic.⁠Source 3, Source 4
Control planeYours to build: Uber built an MCP registry as its control plane and a proxy gateway as its data plane.⁠Source 10In version 2.x, managed by Kong in Konnect, with data plane nodes you run.⁠Source 4
Access controlA2A calls authorization logic implementation-specific.⁠Source 1 MCP makes authorization optional.⁠Source 16API key or OpenID Connect authentication, plus an allow or deny list per agent.⁠Source 9
Agent registryA2A prescribes no standard API for curated registries.⁠Source 14 The official MCP Registry is in preview and does not support private servers.⁠Source 17In 2.x, each AI Agent entity is scoped to one gateway instance.⁠Source 4, Source 9 Kong’s organization-wide agent inventory, in Konnect Catalog, is in beta.⁠Source 15
PricingThe A2A and MCP specifications are openly licensed, A2A under Apache 2.0.⁠Source 1, Source 18 Build and running costs are not publicly documented.AI Management Plus from $25 a month plus usage.⁠Source 19 Enterprise is custom, billed annually.⁠Source 19

What each one is

Build it yourself (DIY)

DIY means connecting and governing agents without buying an agent platform: protocols such as MCP and A2A wired together in-house, existing infrastructure stretched, an in-house platform, self-hosted open source, or no central approach; Uber says that before its gateway, teams built integrations independently, duplicating infrastructure.⁠Source 10

Kong AI Gateway

Kong AI Gateway governs LLM, MCP, and agent-to-agent (A2A) traffic through one data plane, with shared authentication, observability, and policy features.⁠Source 3, Source 4 In 2.x, an AI Agent entity proxies to an agent’s own URL.⁠Source 9

The differences that matter

  1. Agent registry

    DIY

    A2A’s current specification prescribes no standard API for curated registries, and its docs say a curated registry means deploying and maintaining a registry service.⁠Source 14

    Kong AI Gateway

    In 2.x, each agent entity is scoped to one gateway instance; Kong’s organization-wide inventory, Agents in Konnect Catalog, is in beta and not for production.⁠Source 4, Source 9, Source 15

    The official MCP Registry is in preview and does not support private servers; its docs recommend hosting your own registry for those.⁠Source 17

  2. Access control

    DIY

    In A2A, authorization logic is implementation-specific, and MCP makes authorization optional; Pinterest, for example, checks end-user JWTs and mesh identities on almost every MCP call.⁠Source 1, Source 11, Source 16

    Kong AI Gateway

    Kong AI Gateway can require API key or OpenID Connect authentication before routing and enforce a per-agent allow or deny list.⁠Source 9

  3. Where it runs

    DIY

    It runs where you put it: Pinterest optimized for MCP servers in its internal cloud, and Uber’s gateway sends requests through its service mesh sidecar.⁠Source 10, Source 11

    Kong AI Gateway

    Version 2.x is hybrid only, and Kong says its control plane is never in the path of user data traffic.⁠Source 4

For security teams

What a security review asks, answered from each vendor’s public documentation.

TopicDIYKong AI Gateway
Network exposureMCP servers on Streamable HTTP expose an HTTP endpoint, and A2A agents on HTTP use HTTPS URLs in production.⁠Source 1, Source 20In 2.x, nodes you run proxy traffic upstream, including to agents, and authenticate to the control plane over mTLS.⁠Source 4, Source 9
IdentityIn A2A, identity is established at the HTTP layer and credentials are obtained out of band.⁠Source 1, Source 5 MCP authorization is optional.⁠Source 16Can require an API key, or OpenID Connect through Okta, Azure AD, Google, or another OIDC provider, before routing.⁠Source 9, Source 21
Access changes and revocationEach A2A server authorizes requests under its own policies, and the specification calls that logic implementation-specific.⁠Source 1Each agent has an enabled switch and allow or deny list.⁠Source 9 Nodes keep their last configuration if Konnect is unreachable.⁠Source 4
Audit trailA2A docs advise auditing task creation, critical state changes, and agent actions, and tracing, for example with OpenTelemetry.⁠Source 5A2A audit logs record task IDs, method calls, latencies, and errors.⁠Source 22 Konnect audit logs (Enterprise): webhook delivery, kept 7 days.⁠Source 19, Source 23
ComplianceSits with the implementer: A2A docs say to ensure compliance with regulations such as GDPR, CCPA, and HIPAA.⁠Source 5A FIPS 140-3 compliant package, not submitted for NIST validation.⁠Source 24 Kong Inc. can share its SOC 2 report under NDA.⁠Source 25

Full comparison

18 criteria in five groups. Every cell links to its source, or says no public source answers it.

DIY and Kong AI Gateway compared on 18 criteria
DIYKong AI Gateway
What it is
What it is and who it’s forAn in-house build on protocols such as A2A, an open standard for communication between agent systems, and MCP, which the A2A specification calls complementary.⁠Source 1A gateway that governs LLM, MCP, and A2A traffic.⁠Source 3, Source 4 All three run through one data plane, with shared authentication, observability, and policy features.⁠Source 4
MaturityVaries by component: MCP’s latest specification revision is 2026-07-28.⁠Source 2 The official MCP Registry is in preview and may have breaking changes or data resets.⁠Source 17Kong announced AI Gateway in February 2024 and 2.0 general availability in September 2026.⁠Source 26, Source 27 2.2.0 came 30 September 2026.⁠Source 28 Catalog agents are in beta.⁠Source 15
Control
Agent registry and discoveryBuild your own: A2A prescribes no standard API for curated registries, and the official MCP Registry is in preview and does not support private servers.⁠Source 14, Source 17Each AI Agent entity is scoped to one gateway instance.⁠Source 4, Source 9 Konnect Catalog’s agent inventory is in beta.⁠Source 15
Identity and access controlIn A2A, identity is established at the HTTP layer and authorization logic is implementation-specific.⁠Source 1, Source 5 MCP authorization is optional.⁠Source 16API key or OpenID Connect through Okta, Azure AD, Google, or any OIDC provider, then a per-agent allow or deny list.⁠Source 9, Source 21
Ownership, policy, and revocationMCP says implementers should build consent and authorization flows.⁠Source 2 Uber starts every MCP server and tool disabled until reviewed.⁠Source 10Per-agent, per-consumer, or global policies for input validation, logging, and rate limits.⁠Source 4, Source 9 PII sanitizing covers model requests, as an add-on on Plus.⁠Source 19, Source 29
Audit log and observabilityA2A docs advise auditing significant events and tracing, for example with OpenTelemetry.⁠Source 5 MCP documents trace context propagation.⁠Source 30A2A audit logs record task IDs, method calls, latencies, and errors.⁠Source 22 A2A telemetry can go to Konnect and OpenTelemetry.⁠Source 9 Bodies go to Konnect only if you opt in.⁠Source 4
Connection
How agents connectMCP servers on Streamable HTTP expose an HTTP endpoint, and A2A agents on HTTP use HTTPS URLs in production.⁠Source 1, Source 20 AWS PrivateLink privately links a VPC to services.⁠Source 7An AI Agent entity proxies to an upstream agent URL.⁠Source 9 In 2.x, self-managed data plane nodes hold a persistent connection to the Konnect control plane.⁠Source 4
Agents across organizationsA2A is designed for agents from different companies on separate servers.⁠Source 31 Each server authorizes requests under its own policies.⁠Source 1Not publicly documented (checked 2 October 2026)
Protocol supportMCP defines stdio and Streamable HTTP transports.⁠Source 32 A2A maps to JSON-RPC, gRPC, and HTTP/REST bindings.⁠Source 1A2A over JSON-RPC and REST bindings.⁠Source 9 MCP revisions 2025-03-26, 2025-06-18, 2025-11-25, and 2026-07-28 (from version 2.1).⁠Source 33
Frameworks, models, and clouds supportedA2A gives agents built on different frameworks, languages, or vendors a common language.⁠Source 1 Google’s ADK says it is model-agnostic and deployment-agnostic.⁠Source 8Kong says it governs A2A traffic without changing how agents are built.⁠Source 34 Requests to agents that don’t use A2A can pass through as plain HTTP.⁠Source 9
Operations
Deployment options and data residencyWherever you run it: Pinterest optimized for MCP servers in its internal cloud.⁠Source 11 A2A leaves protecting stored data to your own policies.⁠Source 52.x is hybrid: a Konnect control plane, data plane nodes you run.⁠Source 4 Kong’s pricing also lists Kong-managed Dedicated Cloud and serverless AI gateways.⁠Source 19, Source 35, Source 36
Compliance attestationsSits with the implementer: A2A docs cite regulations such as GDPR, CCPA, and HIPAA, and MCP leaves access controls and data protection to implementers.⁠Source 2, Source 5Kong’s trust center lists ISO 27001:2022 and SOC 2 for Kong Inc.⁠Source 25 From 2.2, a FIPS 140-3 compliant package, not submitted for NIST validation.⁠Source 24
Support and SLADepends on the component: MCP SDKs are tiered partly by maintenance commitments.⁠Source 37 The official MCP Registry, in preview, gives no uptime guarantees.⁠Source 17, Source 38Konnect targets 99.9% a month; Dedicated Cloud Gateways list a 99.99% SLA, serverless AI gateways none.⁠Source 19 Enterprise support SLAs: 30 minutes to 2 hours.⁠Source 19
Time and effort to get runningA curated A2A registry is a service you deploy and maintain.⁠Source 14 Pinterest built a unified deployment pipeline after new MCP servers took too much setup.⁠Source 11A quickstart script creates a Konnect control plane and a local Docker data plane.⁠Source 39 To route A2A traffic, you then create an AI Agent entity and attach policies.⁠Source 40
Pricing model and public pricesThe A2A and MCP specifications are openly licensed, A2A under Apache 2.0.⁠Source 1, Source 18 Build and running costs are not publicly documented.Plus: from $25 a month plus usage; control planes $200 hybrid, $500 dedicated, $25 serverless, per month.⁠Source 19 Enterprise: custom, billed annually.⁠Source 19
Building
Agent building toolsFrameworks such as LangGraph and Google’s open-source Agent Development Kit build and deploy agents.⁠Source 8, Source 12 A2A has SDKs in six languages.⁠Source 41Kong says it can generate MCP tools and servers from Kong-managed APIs.⁠Source 3 Tools for building agents are not publicly documented.
Model accessChosen by whoever builds the agents: Google’s ADK says it is optimized for Gemini and model-agnostic.⁠Source 8One API to providers including OpenAI, Anthropic, Amazon Bedrock, Gemini, Azure AI, Mistral, and Ollama, with credentials you supply.⁠Source 4, Source 39, Source 42
Integrations and ecosystemThe official MCP Registry, in preview, has a REST API for clients and aggregators to discover MCP servers.⁠Source 17A Policies Hub of policies and integrations.⁠Source 43 AI Vault works with AWS, GCP, Azure, and HashiCorp Vault secrets backends.⁠Source 4

Which to choose

Choose DIY if

  • You need an organization-wide production registry: Uber and Pinterest run theirs;⁠Source 10, Source 11 Kong’s Catalog Agents is in beta, MCP Registries in tech preview.⁠Source 15, Source 44
  • You already run a service mesh or internal access-control system and want it to check agent calls, as Uber and Pinterest do.⁠Source 10, Source 11
  • You want your own review rules: Uber starts every MCP server and tool disabled until reviewed; Pinterest reviews all but one-off experiments.⁠Source 10, Source 11
  • You want no agent platform contract: the A2A and MCP specifications are openly licensed, A2A under Apache 2.0.⁠Source 1, Source 18

Choose Kong AI Gateway if

  • You want one gateway for LLM, MCP, and A2A traffic, with shared authentication, observability, and policy features.⁠Source 4
  • You want caller authentication and per-agent allow or deny lists without writing them, through API keys or an existing OpenID Connect provider.⁠Source 9, Source 21
  • You want a managed control plane (2.x) outside the traffic path, with nodes you run that keep proxying if it is unreachable.⁠Source 4
  • You want a vendor with published prices, a 30-day free trial on Plus, and Enterprise support up to a 24x7, 30-minute SLA.⁠Source 19

Questions buyers ask

Is Kong AI Gateway an alternative to building it yourself?

For the gateway part, yes: Kong AI Gateway can authenticate callers, apply per-agent access lists and policies, and log A2A calls.⁠Source 9, Source 22 Kong’s organization-wide agent inventory is in beta, and how Kong AI Gateway handles agents owned by other organizations is not publicly documented.⁠Source 15

Does Kong AI Gateway support A2A and MCP?

Yes. It detects A2A requests over JSON-RPC and REST bindings and rewrites agent-card URLs to the gateway address.⁠Source 9 It accepts MCP revisions 2025-03-26, 2025-06-18, and 2025-11-25, and 2026-07-28 from version 2.1.⁠Source 33

Can Kong AI Gateway be fully self-hosted?

Not with the 2.x AI Gateway entities, which Kong’s docs say have no fully self-hosted control plane today.⁠Source 4 Self-hosted Kong Gateway is configured with AI plugins instead, and fully self-hosted AI gateways are part of the separate Gateway Enterprise offering.⁠Source 19, Source 46

How do the costs compare?

On Kong’s AI Management Plus plan, prices start at $25 a month plus usage, and a hybrid AI gateway control plane is $200 a month.⁠Source 19 Enterprise pricing is custom.⁠Source 19 For DIY, the protocol specifications are openly licensed; build and running costs are not publicly documented.⁠Source 1, Source 18

How we compare

Read the full method

Every claim on this page links to a public source. Where none answers a question, the page says so.

We re-check every fact at least every 90 days. This page was last checked .

Something wrong or out of date? Tell us and we’ll correct it.

Sources

51 public sources, each with the date we checked it. Every one opens in a new tab.

  1. Source 1: Agent2Agent (A2A) Protocol Specification a2a-protocol.org · checked Back:abcdefghijklmnopq

  2. Source 2: Specification - Model Context Protocol 2026-07-28 modelcontextprotocol.io · checked Back:abcd

  3. Source 3: Kong AI Gateway product page Kong · checked Back:abcdefg

  4. Source 4: AI Gateway architecture - Kong Docs Kong · checked Back:abcdefghijklmnopqrstuv

  5. Source 5: Enterprise Features - A2A Protocol a2a-protocol.org · checked Back:abcdefghi

  6. Source 6: Security - Istio concepts istio.io · checked Back to text

  7. Source 7: What is AWS PrivateLink? - Amazon Virtual Private Cloud docs.aws.amazon.com · checked Back:ab

  8. Source 8: google/adk-python README (GitHub) github.com · checked Back:abcde

  9. Source 9: AI Agents - Kong AI Gateway docs Kong · checked Back:abcdefghijklmnopqrst

  10. Source 10: Designing MCP Gateway Uber's MCP Management Platform - Uber Blog uber.com · checked Back:abcdefghi

  11. Source 11: Building an MCP Ecosystem at Pinterest - Pinterest Engineering Blog medium.com · checked Back:abcdefgh

  12. Source 12: langchain-ai/langgraph README (GitHub) github.com · checked Back:ab

  13. Source 13: Integrating with Model Context Protocol (MCP) - Keycloak keycloak.org · checked Back to text

  14. Source 14: Agent Discovery - A2A Protocol a2a-protocol.org · checked Back:abcde

  15. Source 15: Agents in Catalog - Kong Docs Kong · checked Back:abcdefg

  16. Source 16: Authorization - Model Context Protocol specification 2026-07-28 modelcontextprotocol.io · checked Back:abcd

  17. Source 17: The MCP Registry - Model Context Protocol modelcontextprotocol.io · checked Back:abcdef

  18. Source 18: modelcontextprotocol/modelcontextprotocol LICENSE (GitHub) github.com · checked Back:abcd

  19. Source 19: Kong Pricing & Plans Kong · checked Back:abcdefghijklm

  20. Source 20: Streamable HTTP - Model Context Protocol specification 2026-07-28 modelcontextprotocol.io · checked Back:ab

  21. Source 21: AI Auth Strategies - Kong AI Gateway docs Kong · checked Back:abcd

  22. Source 22: Route A2A traffic through AI Gateway - Kong Docs Kong · checked Back:abc

  23. Source 23: Konnect and Dev Portal audit logs - Kong Docs Kong · checked Back to text

  24. Source 24: FIPS 140-3 compliance in AI Gateway - Kong Docs Kong · checked Back:ab

  25. Source 25: Trust Center - Kong Inc. Kong · checked Back:ab

  26. Source 26: Announcing Kong’s New Open Source AI Gateway with Multi-LLM Support, No-Code AI Plugins, Advanced Prompt Engineering, and More Kong · checked Back to text

  27. Source 27: Kong AI Gateway 2.0 Is Now GA - And It's Already Moving Faster Kong · checked Back to text

  28. Source 28: Kong AI Gateway changelog - Kong Docs Kong · checked Back to text

  29. Source 29: AI PII Sanitizer - Policy | Kong Docs Kong · checked Back to text

  30. Source 30: Key Changes - Model Context Protocol specification 2026-07-28 modelcontextprotocol.io · checked Back to text

  31. Source 31: a2aproject/A2A README (GitHub) github.com · checked Back to text

  32. Source 32: Transports - Model Context Protocol specification 2026-07-28 modelcontextprotocol.io · checked Back to text

  33. Source 33: MCP version support - Kong AI Gateway docs Kong · checked Back:ab

  34. Source 34: The Agent Gateway for Secure, Observable Agent-to-Agent Communication (Kong) Kong · checked Back:ab

  35. Source 35: Dedicated Cloud Gateways - Kong Docs Kong · checked Back to text

  36. Source 36: Serverless Gateways - Kong Docs Kong · checked Back to text

  37. Source 37: SDK Tiers - Model Context Protocol modelcontextprotocol.io · checked Back to text

  38. Source 38: MCP Registry Aggregators - Model Context Protocol modelcontextprotocol.io · checked Back to text

  39. Source 39: Kong AI Gateway | Kong Docs Kong · checked Back:ab

  40. Source 40: Route A2A agent traffic through AI Gateway - Kong Docs Kong · checked Back to text

  41. Source 41: A2A protocol roadmap a2a-protocol.org · checked Back to text

  42. Source 42: AI Gateway providers - Kong Docs Kong · checked Back to text

  43. Source 43: Kong AI Gateway Policies Hub - Kong Docs Kong · checked Back to text

  44. Source 44: MCP registries in Catalog (tech preview) - Kong Docs Kong · checked Back to text

  45. Source 45: AI Gateway audit log reference - Kong Docs Kong · checked Back to text

  46. Source 46: Configure Kong AI Gateway on-prem - Kong Docs Kong · checked Back to text

  47. Source 47: Your company's private network Blocks.ai · checked Back to text

  48. Source 48: Network requirements Blocks.ai · checked Back to text

  49. Source 49: Solutions: Agent sprawl Blocks.ai · checked Back to text

  50. Source 50: Solutions: Partner networks Blocks.ai · checked Back to text

  51. Source 51: Pricing Blocks.ai · checked Back to text