How to Choose an MCP Provider: 9 Questions to Ask

What an MCP provider does

An MCP provider runs MCP servers so you do not have to. You authorize a tool such as Shopify or HubSpot once, and the provider gives your AI client a URL to call. You will also see the terms hosted MCP server, managed MCP and MCP gateway. They overlap, and the questions below apply to all three.

If the protocol is new to you, read what MCP is first. If you only need one tool on one account, the vendor's own server may be enough. The best MCP servers lists those.

The 9 questions

1. Where are my credentials stored, and how do I cut a client off?

The provider holds the keys to your store, ad accounts or books. Two things should be written down somewhere public: how those credentials are encrypted, and that the AI client never receives them. The MCP security guidance forbids token passthrough, which means a server must not accept or forward a token that was not issued for it.

Then look for the off switch. Stripe's dashboard, for example, lists each MCP client's OAuth session with a control to end it. Ask any provider where its equivalent is.

Walk away if the setup guide tells you to paste a provider API key into the AI client's config file.

2. Do I get the full API or a curated tool set?

Some servers ship a short list of tools for the common tasks. That works until you need the endpoint they left out.

Look for a public tool list per integration, readable before you pay. Then look for a way to reach the rest of the API without loading thousands of tool definitions into the model's context. Stripe does it with two generic tools, stripe_api_read and stripe_api_write. Others use a search tool. MCP token limits explains why this matters.

"We support HubSpot" with no tool list behind it is not an answer.

3. Is pricing metered per call or flat?

You cannot predict how many tool calls a conversation will make. One question can trigger several. So check whether calls are the billing unit, and if they are, whether the rate is public.

Two examples of metering you can model. Zapier counts two tasks for each successful MCP tool call. Smithery's paid plan includes a $10 monthly credit, then charges $0.10 per 1,000 RPCs. Flat plans priced on connections, seats or workspaces are easier still to forecast.

The thing to avoid is a "fair use" limit or a credit with no stated conversion into calls.

4. Can I connect more than one account per tool?

Agencies, multi-brand retailers and anyone with a staging store need this. Some integrations assume one account. Shopify's official AI integrations connect one store at a time, and switching disconnects the current store and starts a new authorization.

Look for separate workspaces or environments, each with its own connections and usage data. Multiple accounts, one MCP and MCP for agencies go deeper.

5. Which AI clients does it support?

Your team may use Claude, ChatGPT and Cursor in the same week. A provider should offer a standard remote endpoint: Streamable HTTP with OAuth sign-in. Check its client list against the clients you actually use, and ask about the older SSE transport if one of yours still needs it.

A provider that needs a local process on every laptop is a different kind of product. See remote vs local MCP servers.

6. Can I make access read-only, and can a human approve writes?

These are two separate controls.

Read-only means the write and delete tools are not exposed at all, per person or per connection. Telling the model not to write is only a prompt. It does not stop a tool call.

Approval means a risky write waits for a person. Stripe's server requires confirmation for refunds and outbound payments. Other providers leave approval to the AI client. Claude shows a tool approval request unless you chose "Allow always" for that tool. Know which layer you are relying on.

7. What gets logged, and for how long?

When something changes in a client's account, you need to see which tool ran, when, who triggered it and from which AI client. Check the retention period on the plan you would buy, because it can differ by tier.

Ask whether the logs hold payloads. Providers differ on this. Glama's hosting page says its gateway logs every JSON-RPC call with full payloads. PopMCP stores metadata only. Full payloads help an audit and put your customers' data in the provider's logs. Metadata keeps that data out, and you lose the record of exactly what the model read.

8. What compliance paperwork exists?

A larger client's security team will ask before you connect their accounts. Look for a published DPA, a subprocessor list, a stated data location, SSO, and an audit report such as SOC 2 if your buyers require one.

Ask for the report itself. A badge on a homepage proves little, and a claim you cannot read should count as absent.

9. What happens when the provider's API changes?

APIs under an MCP server move. Shopify removed the catalog and cart tools from its Storefront MCP endpoint in favour of a newer protocol. Stripe's MCP server stops accepting some API key types on 31 October 2026.

A provider should own those upgrades, keep tool names stable and warn you before removing a tool. The evidence is a public changelog. HubSpot announced its MCP server's general availability in its developer changelog, and Stripe put a dated notice in its MCP docs.

A community server with one maintainer and no release this year gives you none of that.

How PopMCP answers the 9 questions

We run PopMCP, so here are our own answers. Five of the nine have a weak spot, and each is marked.

If a missing SOC 2 report, self-serve SSO or an approval step rules us out, that is a fair call. You can compare other hosted options on our Composio, Zapier MCP and MintMCP pages. The plan details behind rows 3 to 7 are on the pricing page and in the FAQ.