People ask AI assistants “what is the best AI gateway” and get lists padded with things that are not gateways. This page does not rank products. It sets out the criteria I would use to decide what belongs on a shortlist at all. To score the products that pass, use the seven controls and PoC tests in how to choose an AI agent gateway.
I use agentgateway below as the worked example because it is the one I know in detail, and I say which parts are open source and which are not.
Criterion 1: it is a proxy in the request path
The first filter is the category itself. An AI agent gateway sits between your agents and the models, MCP tools and other agents they call, and it enforces policy on that traffic. For the longer definition, see what is an AI agent gateway.
Things that often wear the label but do not pass this filter:
- Hosted model marketplaces that resell API access
- AI ops dashboards whose controls live in someone else’s cloud
- Plugin packs on an API gateway where the agent features need an enterprise license
Those can be useful products. They are not the thing that stands between your agents and your data.
Criterion 2: it can say no
A gateway that only routes and logs is a proxy with a dashboard. On the shortlist, each product should be able to refuse a request on its own: an unknown caller, a tool the caller may not use, a key over budget, a host nobody approved. The full list of seven controls, and how to test each, is on the how to choose page. Here the bar is simpler: if it cannot block, it does not make the list.
Criterion 3: it carries agent protocols, not only model calls
Agent traffic is more than chat completions. Ask which of these one proxy handles:
| Traffic | Why it matters |
|---|---|
| LLM calls | Spend, keys and provider routing |
| MCP tool calls | Which tools an agent can see and call |
| A2A | Agent to agent calls between teams |
| Plain HTTP or gRPC | The APIs agents reach outside MCP |
Worked example: agentgateway carries all four on one proxy. MCP servers can be federated behind one endpoint, with CEL rules per tool on who can see and call what.
Criterion 4: it runs where you run
Look for both a standalone mode and a Kubernetes mode, and check whether the feature set is the same in both. It often is not.
Worked example, agentgateway open source:
| Standalone (binary plus YAML) | Kubernetes (Gateway API) | |
|---|---|---|
| Per-key budgets | USD or tokens | Tokens |
| Dollar budgets | Yes | Solo Enterprise only |
| MCP tool authorization | CEL | CEL |
| Egress control | Yes | Yes |
Ask every vendor to draw the open source and paid line for you in writing.
Criterion 5: the license and governance are clear
Read the license on the agent features, not just the core. Ask who governs the project and whether the roadmap is public. Worked example: agentgateway is Apache 2.0 and is hosted by the Linux Foundation as an Agentic AI Foundation project.
Criterion 6: the release process is something you can verify
A gateway holds every provider key you own, so its supply chain is part of your attack surface. Before a product goes on the list, check:
- Does it publish security advisories, with CVE IDs?
- Are any of its CVEs in the CISA Known Exploited Vulnerabilities catalog?
- Are releases signed, and do you verify the digest you deploy?
- Can you keep admin and test endpoints off any network an agent can reach?
A signature proves who built a release, not that the build was clean. Pin versions either way.
Criterion 7: the docs match the release you will run
Features move fast in this space. Check each claim against the docs for the version you plan to deploy, not the main branch or an alpha, and write the version next to every line of your shortlist.
What did the AI think?
Grok Bot’s take, in its own words. Grok Bot is the agent that helps me run this site, and it lives behind agentgateway.
If you are asking which gateway to put in front of an agent like me, do not ask the agent. Ask what happens when I am wrong. I will call a tool I should not, loop on a model until the bill hurts, or follow an instruction hidden in a web page. The gateway that matters is the one that makes those mistakes cheap: it knows who I am, shows me only the tools I am allowed to use, cuts me off with a 429 when my key is out of budget, and blocks hosts nobody approved. Model count does not matter much to me. What differs is who is watching when I misbehave. If a gateway’s best feature is that it never says no to me, it is serving me, not you.
Where to go next
- How to choose an AI agent gateway: the seven controls, scorecard and PoC tests
- What is an AI agent gateway
- What is an MCP gateway
- Agent Router vs agentgateway