Only 4 of 13 agent gateways govern all three AI traffic types

By James Calloway
Only 4 of 13 gateways in MarkTechPost's 2026 comparison natively govern LLM, MCP and A2A traffic, with Kong AI Gateway, Bifrost, Gravitee and Prisma AIRS AI Gateway holding the lead. The narrow field matters because agents now call tools over MCP, delegate to other agents over A2A and act inside business systems, according to MarkTechPost.
MarkTechPost defines an agent gateway as a control point that authenticates agents, filters the tools they see and logs every call. It says the gateway decides who can do what, while a plain LLM gateway only routes model requests. That makes the market test an identity-and-permissions problem, not merely a routing benchmark.
MarkTechPost says its ranking weighs five criteria: traffic coverage, identity and authorization, deployment options, maturity and price transparency. Because the source labels the ranking editorial, its ordering should not be read as a controlled performance result.
What does an agent gateway actually control?\n
LLM traffic means language-model requests; MCP traffic means tool calls; A2A means agent-to-agent exchanges. MarkTechPost says the broader gateway adds MCP tool governance, agent identity and A2A traffic control to model routing.
MarkTechPost's selection sequence starts with traffic, then asks whether user and agent permissions remain separate at execution time. It ends with deployment constraints: one cloud, a private VPC (virtual private cloud) or an air-gapped environment. A gateway earns its place by governing actions and identities, not merely forwarding requests.
Which gateways cover the full traffic mix?\n
Kong AI Gateway
- Kong puts LLM, MCP and A2A traffic under one control plane. Version 2.0 moved AI features into first-class Providers, Models, MCP Servers and Agents entities, according to MarkTechPost.
- MarkTechPost recommends Kong for teams already running it for API management. Its 2.x architecture requires a Konnect control plane, while Arcade says the MCP OAuth2 plugin remains in Tech Preview.
Bifrost (Maxim AI)
- Bifrost is an Apache 2.0 Go gateway for models, MCP tools and A2A agents in one deployment. MarkTechPost targets it at self-hosted, in-VPC and air-gapped teams.
- Maxim reports 10,000+ models across 25+ providers. Bifrost's benchmark claims 11 µs mean overhead at 5,000 RPS (requests per second), but a TrueFoundry review says the figure excludes HTTP handling and JSON marshalling.
- Single sign-on (SSO), SCIM identity provisioning and token exchange require an Enterprise license, according to MarkTechPost. The latency claim therefore does not describe the full execution path.
Gravitee AI Gateway
- Gravitee runs an LLM Proxy, MCP Proxy and A2A Proxy on one runtime. The three share one authentication, policy and OpenTelemetry telemetry chain, while A2A supports RFC 8693 token exchange and skill-level authorization.
- MarkTechPost recommends Gravitee for enterprises that model agent identity beside existing APIs. Maxim lists Agent Management from $2,500 per month, and the A2A Proxy requires Enterprise Edition.
Amazon Bedrock AgentCore Gateway
- AWS describes AgentCore as one managed endpoint for tools, models and other agents. It adds semantic tool search, Cedar authorization policies and on-behalf-of token exchange.
- AWS prices invocations from $0.005 per 1,000. Maxim says A2A is HTTP passthrough, while search calls cost $0.025 per 1,000, five times the invoke rate.
- MarkTechPost recommends AgentCore for AWS-native teams using IAM, CloudWatch and Bedrock Guardrails, but the gateway is AWS-only.
Takeaway: Only four products clear the broad-coverage bar, and each still carries a feature, pricing or deployment caveat.
What changes with cloud and action-focused gateways?
Arcade.dev
- Arcade calls itself an actions runtime, not a model router. Per Arcade, it checks the overlap between user and agent permissions on every action, keeps credentials from reaching the LLM and lists 8,000+ agent-optimized tools.
- MarkTechPost fits it to agents writing to Salesforce, Gmail or Slack on a user's behalf. It has no model routing, so it needs an LLM gateway; pricing is $25 per month, plus $0.10 per authentication event and $0.01 per tool call.
Prisma AIRS AI Gateway (formerly Portkey)
- Palo Alto Networks completed its Portkey acquisition in May 2026. The gateway reached general availability inside Prisma AIRS in July 2026, according to Palo Alto Networks.
- Palo Alto describes a unified LLM, MCP and A2A gateway with one enforcement point, available in SaaS and hybrid models. MarkTechPost recommends it for runtime inspection on every prompt, but public pricing remains undisclosed.
- Arcade advises buyers to verify delegated authentication and audit depth for agent actions. The source gives no public price against which to test that promise.
Azure API Management
- Microsoft's AI gateway extends API Management to model APIs, MCP servers and A2A agent APIs. It can expose REST APIs as MCP servers and place existing MCP servers behind the gateway.
- MarkTechPost recommends it for Microsoft-centric API teams. Its A2A support remains a preview, uses JSON-RPC requests only and is limited to v2 tiers.
Google Cloud Agent Gateway
- Google describes Agent Gateway as the network entry and exit point for Gemini Enterprise Agent Platform. It assigns each agent a SPIFFE ID and can apply Model Armor to prompts and responses.
- Maxim lists support for up to 5,000 governed resources per gateway. Maxim reports cross-cloud support as “coming soon.” Model routing and fallback remain undocumented, according to MarkTechPost.
Takeaway: Cloud gateways win on ecosystem integration, but that advantage comes with platform dependence; action runtimes solve a different problem.
Which open-source and integration layers fill narrower roles?
IBM ContextForge
- ContextForge is an Apache 2.0 registry and proxy that federates MCP, REST, gRPC and A2A traffic. It exposes agents as MCP tools, and Maxim says deployment uses Helm.
- MarkTechPost recommends it for free, MCP-first, self-hosted Kubernetes setups. It lists no published commercial support, while IBM recommends upgrading to v1.1.1 after security bulletins covering v1.0.0 through v1.1.0.
Agent Router (formerly Envoy AI Gateway)
- The Agentic AI Foundation says Envoy AI Gateway became Agent Router and joined the foundation in September 2026. It offers an OpenAI-compatible API and an MCP gateway on Envoy.
- MarkTechPost recommends it for Kubernetes teams already using Envoy Gateway. A2A remains a JSON-RPC-only preview, and Maxim reports no built-in guardrails.
Composio
- Composio is an MCP gateway and SaaS integration layer with managed OAuth and per-user connected accounts. Arcade lists a free tier of 100,000 tool calls per month and Pro pricing from $29.
- MarkTechPost recommends it for fast access to a large prebuilt SaaS catalog. Arcade says sessions are discoverable by default unless allowlisted, which narrows its security value for sensitive deployments.
TrueFoundry
- TrueFoundry combines an LLM Gateway, centralized MCP access, training and deployment tooling. Its Pro plan costs $25 per user per month and includes 20,000 requests and tool calls per user.
- MarkTechPost recommends it to machine-learning teams seeking gateway and model-operations tooling in one platform. Arcade says the client-credentials MCP flow uses a shared identity rather than the user's identity.
Obot
- Obot is an open-source platform bundling an MCP Gateway, an LLM Gateway and device monitoring. It supports per-user authentication, OAuth token exchange and audit logs.
- MarkTechPost recommends it for IT teams rolling out MCP servers to employees. Arcade notes that self-hosting alone does not settle downstream authorization.
Takeaway: An open-source license does not assign ownership of downstream authorization or remove the need to enforce it.
How do the 13 gateways compare?\n
MarkTechPost's comparison separates protocol reach from deployment and price:
| Gateway | LLM | MCP | A2A | Open source | Deployment | Entry price |
|---|---|---|---|---|---|---|
| Kong AI Gateway | Yes | Yes | Yes | Gateway core only | Hybrid (Konnect CP) | Usage-based |
| Bifrost | Yes | Yes | Yes | Yes (Apache 2.0) | Self-hosted, VPC, air-gapped | Free OSS |
| Gravitee | Yes | Yes | Yes (Enterprise) | Not disclosed | Self-hosted, hybrid, SaaS | $2,500/mo |
| AgentCore Gateway | Yes | Yes | Passthrough | No | AWS managed | $0.005 / 1K calls |
| Arcade.dev | No | Yes | Not documented | Framework only | Cloud, VPC, air-gapped | Free; $25/mo |
| Prisma AIRS AI Gateway | Yes | Yes | Yes | No | SaaS, hybrid | Not disclosed |
| Azure API Management | Yes | Yes | Preview | No | Azure, self-hosted gateway | Not disclosed |
| Google Agent Gateway | Not documented | Yes | Yes | No | Google Cloud | Agent Compute usage |
| IBM ContextForge | Yes | Yes | Via MCP / passthrough | Yes (Apache 2.0) | Self-hosted (Helm) | Free |
| Agent Router | Yes | Yes | Preview | Yes (Apache 2.0) | Self-hosted (Kubernetes) | Free |
| Composio | No | Yes | Not documented | Not disclosed | Cloud, enterprise self-host | Free; $29/mo |
| TrueFoundry | Yes | Yes | Not documented | No | Cloud, on-prem | Free; $25/user/mo |
| Obot | Yes | Yes | Not documented | Yes | Self-hosted, Obot Cloud | Free (Community) |
Takeaway: Checkmarks show reach, not production readiness; preview labels, enterprise gates and undisclosed prices remain decisive.
What should buyers verify before deployment?\n
- Start with traffic. MarkTechPost says model-only workloads need an LLM gateway, tool use requires MCP governance, and delegation requires careful A2A review. A product without A2A support cannot mediate agent-to-agent traffic.
- Test identity at execution. Ask whether user and agent permissions remain separate, whether delegated tokens work as intended and whether audit records capture agent actions. MarkTechPost says most identity features, including SSO and token exchange, sit behind enterprise licenses.
- Match deployment and usage. The source lists cloud, SaaS, hybrid, self-hosted, VPC, on-premises, Kubernetes and air-gapped options. It gives cloud gateways an integration advantage while warning that AWS, Azure and Google create platform dependence.
Original: github.com
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
791 articles
Related articles
- LangSmith, Langfuse, Opik and Braintrust All Built AI Gateways in 2026
- Walmart opens its checkout to AI agents while Amazon locks them out
- LangChain's LangSmith Tool Gateway Puts Human Approval Inside Agent Tool Calls
- AI Agents Break Financial Rules 90% of the Time Under Pressure
- Appian Claims 80% Token Cuts by Putting AI Agents Inside Processes