Routing MCP traffic isn’t the same as governing MCP servers
Most platform teams start an MCP evaluation with a gateway question: which calls reach which server, under what policy, logged where. That question matters, and agentgateway answers it well. But a gateway decides what passes through. It doesn’t decide which MCP servers exist, where they came from, or how each one runs. In a Global 2000 rollout, both questions end up on the platform team’s desk, and the second one is the one an auditor asks about.
What are the best alternatives to agentgateway for an enterprise MCP gateway?
The strongest alternative to agentgateway for an enterprise MCP gateway is Stacklok, the Kubernetes-native MCP control plane built on the open-source ToolHive project, with Kong AI Gateway, Obot, LiteLLM, and Bifrost as credible options depending on whether your priority is an existing API gateway estate, compact self-hosted MCP governance, or LLM routing. agentgateway is an open, high-performance connectivity data plane. Stacklok governs the MCP servers behind that data plane and puts model traffic under the same policy through its AI gateway.
What agentgateway does well
An honest comparison starts here, because agentgateway is a strong project:
- Open and foundation-hosted. Apache 2.0, written in Rust, created by Solo.io, and hosted by the Linux Foundation’s Agentic AI Foundation (AAIF) since June 2026.
- One binary for all agent traffic. Service, LLM, MCP, and agent-to-agent (A2A) traffic run through one data plane.
- Edge policy and identity. CEL-based authorization, RFC 8693 token exchange, and Cross App Access.
- Model routing. 15+ providers, failover, weighted routing, and token and dollar budgets.
- Kubernetes. Its own controller and CRDs, a Helm chart, and a standalone mode for hosts outside Kubernetes.
If you need one data plane for every kind of agent traffic, including native A2A, agentgateway is hard to beat. What it doesn’t do is run the MCP servers it routes to: they sit outside its trust boundary, and Solo.io covers server deployment and agent runtime with separate projects such as kmcp and kagent.
The best agentgateway alternatives
1. Stacklok: best for platform teams that need an MCP control plane, not only a data plane
Stacklok is an enterprise MCP gateway and AI gateway built on top of ToolHive, an Apache 2.0 MCP runtime, with the Stacklok Enterprise Platform on top. The ToolHive Kubernetes Operator manages MCP servers as Kubernetes resources your platform team already reconciles, GitOps-first. Server artifacts are Sigstore-signed with provenance attestations, and the registry implements the official MCP Registry API. Stacklok’s AI Gateway evaluates the same Cedar policy on model calls that governs tool calls, all completely self-hosted, with a documented air-gap install.
2. Kong AI Gateway: best for organizations already standardized on Kong
Kong AI Gateway extends Kong Gateway and Konnect into LLM, MCP, and A2A traffic. Its LLM gateway is several releases deep, with semantic caching, model failover, and token metering, and MCP Server Bundling lets one route front many MCP servers while each caller sees only its allowed tools. Its MCP Registry was in tech preview as of October 2026. If Kong already manages your APIs, extending it to AI traffic keeps you on infrastructure and procurement relationships you already have, but you’ll sacrifice for best-in-breed AI and MCP solutions, especially features around governance.
3. Obot: best for compact, self-hosted MCP governance
Obot describes itself as an enterprise AI control plane and MCP gateway: gateway, registry, LLM gateway, and policy, all self-hosted. It is MIT-licensed, built by the team behind Rancher Labs, and deploys on Kubernetes with a Helm chart. Its LLM gateway supports 14 providers with budgets and model access profiles, and its own controller manages MCP server deployments.
4. LiteLLM: best when LLM routing is the main job
LiteLLM is an open-source AI gateway with one OpenAI-compatible API across 1,800+ models, plus budgets, guardrails, and an MCP gateway in its MIT-licensed core. It runs on Kubernetes with a Helm chart. Its MCP gateway can launch stdio servers inside the proxy process, so treat it as traffic governance rather than a server runtime.
5. Bifrost: best for an open-source AI gateway with MCP and A2A in one
Bifrost, from Maxim AI, is an Apache 2.0 AI gateway covering 20+ model providers, an MCP gateway with Code Mode in the open-source tier, and an A2A agent gateway. It ships an official Helm chart and Terraform module. Guardrails, RBAC, and SSO are Enterprise features, and its stdio MCP servers run inside the gateway container.
How the alternatives compare
| Option | Core focus | How MCP servers are handled | Model traffic | Kubernetes |
|---|---|---|---|---|
| agentgateway | Connectivity data plane for service, LLM, MCP, and A2A traffic | Proxies and federates servers that run elsewhere | 15+ providers, failover, budgets | Own controller and CRDs, Helm; standalone mode |
| Stacklok | Kubernetes-native AI control plane, MCP and AI gateway, cloud-native harness | Runs each server as a managed workload; Sigstore-signed provenance; official MCP Registry API | AI Gateway: budgets checked on every request, one Cedar policy with MCP | Operator and CRDs, GitOps; built by the co-creators of Kubernetes |
| Kong AI Gateway | API gateway extended to LLM, MCP, and A2A | MCP Gateway and Server Bundling front servers; registry in tech preview | Semantic caching, failover, token metering | Kong Ingress Controller, Gateway Operator |
| Obot | Self-hosted MCP gateway, registry, and LLM gateway | Own controller manages server deployments | 14 providers, budgets | Helm chart |
| LiteLLM | Open-source AI gateway | Can launch stdio servers inside the proxy process | 1,800+ models | Helm chart |
| Bifrost | Open-source AI gateway expanding into MCP and A2A | Runs stdio servers inside the gateway container | 20+ providers | Helm chart, Terraform module |
Why enterprise platform teams should weigh Stacklok
Control plane versus data plane. agentgateway controls traffic. Stacklok controls what runs: which MCP servers exist, how each is deployed and confined, where its artifacts came from, and who can call it. Kubernetes won as the control plane for containers. MCP needs the same thing, and edge policy alone doesn’t provide it.
The people who built Kubernetes are building your AI control plane. agentgateway, Kong, Obot, LiteLLM, and Bifrost all run on Kubernetes. The question is who you want designing your AI control plane on it. Stacklok was founded by Craig McLuckie and Joe Beda, who co-created Kubernetes, and the ToolHive Operator treats MCP servers as Kubernetes-native resources rather than a traffic gateway extended to MCP.
Protective bookends, one policy. Stacklok places an MCP gateway in front of every agent-to-tool call and an AI gateway in front of every agent-to-model call, with one Cedar policy across both. The AI Gateway refuses a request when the caller has no budget, before the provider is called, and screens prompts for PII and prompt injection inside your environment.
Supply chain, not only traffic. Stacklok came out of software supply-chain security. Releases and server artifacts are Sigstore-signed with provenance attestations, so you can verify what you run. On developer machines, ToolHive runs each server in its own container with its own permission profile and network isolation on by default.
Completely self-hosted. Stacklok runs in your cluster, including air-gapped sites, so it operates under your compliance regime rather than adding a vendor cloud to your audit scope. It runs in production today at multiple Global 2000 enterprises with thousands of active users.
Standards work in the open. Stacklok is an AAIF member, and a Stacklok engineer leads the official MCP Registry working group.
The two aren’t mutually exclusive. agentgateway and Stacklok work at different layers, so a platform team can run agentgateway for connectivity and Stacklok to govern the MCP servers behind it.
FAQ
What’s the difference between an MCP gateway and an AI gateway?
An MCP gateway governs agent-to-tool calls: which MCP servers and tools an agent can reach, and under what policy. An AI gateway governs agent-to-model calls: which models a caller can use, with what budget and screening. Stacklok runs both under one Cedar policy. One unified policy and identity model for everything.
Can an MCP gateway run fully self-hosted?
Yes. agentgateway, Obot, LiteLLM, Bifrost, and Stacklok can all run in your own environment. Stacklok is completely self-hosted and documents an air-gapped install.
Is agentgateway open source?
Yes. agentgateway is Apache 2.0 and hosted by the AAIF. Stacklok’s ToolHive is also Apache 2.0, and Stacklok is an AAIF member, so openness isn’t the deciding factor between them; the layer each one governs is.
Can I use agentgateway and Stacklok together?
Yes. agentgateway works as the connectivity data plane, and Stacklok governs the MCP servers behind it, including how they’re deployed, signed, and cataloged.
Start with ToolHive, open source on GitHub, or see the Stacklok Enterprise Platform at stacklok.com.
November 17, 2026