How to Choose an MCP Gateway: 12 Questions to Ask

Choosing an MCP gateway is choosing who holds the keys to every tool your agents touch, so the evaluation deserves better than a feature-grid skim. Here are the 12 questions we would ask any vendor, ours included, sorted into security, cost, architecture, and the one question that is strangely absent from every other checklist. For each, what a good answer sounds like.

If you have not settled whether you need a gateway at all, [start there](/blog/do-you-need-an-mcp-gateway); this post assumes the answer was yes.

Security: questions 1 to 4

1. How are credentials stored? Good answer: encrypted per secret, decrypted only at call time, never visible to the agent or in URLs. Bad answer: anything vague, or a shared master key wrapping everyone's secrets. This is the heart of [gateway security](/blog/mcp-gateway-security), and vagueness here ends the evaluation.

2. Can I revoke one connection without touching the rest? Good answer: one click, immediate, per connection. The test case is a departing contractor or a misbehaving integration: seconds matter, and re-authenticating everything else as a side effect means you will hesitate when it counts.

3. Is there a per-call audit trail? Good answer: every routed call logged with tool, arguments, result, initiator, timestamp, and you can filter it. "We have logs" is not the same as [an audit trail you can actually query](/blog/ai-agent-audit-trail). Also ask who can see it and how long it is kept.

4. How are OAuth scopes handled? Good answer: scopes requested per connection, visible afterwards, minimal by default. A gateway that requests broad access everywhere "so things work" has moved your risk to its convenience.

Cost: questions 5 to 7

5. What exactly is metered? Per tool call, per task, per seat, per gateway: the models differ by orders of magnitude at the same usage. We published [a verified pricing survey](/blog/mcp-gateway-pricing) because half the market hides numbers; make any vendor state theirs in writing.

6. What happens on failure and on utility calls? Good answer: failed calls are not billed (ours are refunded), and utility calls the agent should make freely (memory, discovery) are free. Metering those quietly taxes good agent behavior.

7. Is there a real free tier and public pricing? A vendor without public pricing is quoting you a sales process, not a product. Fine for enterprises with procurement teams; disqualifying if you are one person with a card.

Architecture: questions 8 to 11

8. Hosted or self-hosted, and where? Both are legitimate; the answer must match your constraints. If EU data residency matters to you, ask for regions per component (processing, storage, logs, memory), not a slogan; [why EU teams ask](/blog/eu-hosting-for-ai-infrastructure) covers the specifics. Self-hosters: the [open source field](/blog/open-source-mcp-gateways) is genuinely strong in 2026.

9. How does federation actually work? Good answer: namespaced tools (service__tool), runtime catalog updates when connections change, upstream auth brokered per server. If the vendor cannot explain [federation](/blog/what-is-mcp-federation) crisply, they are selling a proxy with a thesaurus.

10. What does it do about tool-schema bloat? Every connected server dumps schemas into context by default; a gateway is the one place that can control what loads. Ask concretely: if I connect ten servers with 150 tools, what does my agent's context carry per turn?

11. Does it track the spec? MCP v2 went final in July 2026. Good answer: speaks the current spec, keeps compatibility with older clients, and can front mixed server generations. A gateway that lags the protocol becomes the bottleneck it was meant to remove.

The question nobody asks: number 12

12. Does it remember anything between sessions, or is it a stateless pipe?

Every other checklist stops at routing and governance. But if your agent starts every session from zero, the gateway watched your context get rebuilt at full price and did nothing, and that rebuild is one of the two structural costs in [your token bill](/blog/where-your-llm-token-bill-comes-from). Ask: can the agent store a decision today and recall it next week, from a different client? Is memory scoped per project or client? Can I export and delete it? Are memory calls metered?

We checked the 2026 market while researching our comparison articles, and as of late July, no other gateway we verified describes cross-session memory. So the realistic answers you will hear are "no" and, from us, "yes, and here are the limits". [What changes when the gateway remembers](/blog/mcp-gateway-with-memory) is the long version.

The scorecard, and our own answers

Take the 12 questions into every call and demand written answers. Weight security highest (a gateway with weak answers on 1 to 3 is disqualified at any price), cost by your actual volume, architecture by your constraints, and question 12 by how much of your bill is re-established context.

Since we insist everyone answers, here is [Tulimoa's](/gateway) sheet in one honest paragraph: per-secret encrypted vault, one-click per-connection revocation, full per-call audit trail, scoped OAuth. Public pricing with a free tier, failed calls refunded, memory calls never metered. Hosted in the EU (Frankfurt), namespaced federation on the current spec. Memory: yes, scoped, exportable. Limits: early beta, no self-serve signup yet, one connected agent per plan today, no certifications to wave. Judge that against the same 12 questions you put to everyone else.

Frequently asked questions

What is the most important criterion when choosing an MCP gateway?

Credential handling. Everything else is improvable later; a gateway that stores secrets carelessly is a liability from day one. Encrypted per-secret storage, call-time-only decryption, and instant per-connection revocation are the floor.

Should a small team choose hosted or self-hosted?

Hosted, usually: the ops burden of running credential infrastructure exceeds most small teams' capacity. Self-host when you have platform engineers and a compliance mandate that requires it.

How do I compare gateways with different pricing models?

Estimate your monthly tool calls, compute each offer at that volume, and add the cost of what is not included (audit, vault features, seats). Per-call and per-task models differ by orders of magnitude at identical usage, so never compare stickers.