Connecting an MCP server means more than granting it tools. Every argument you send and every result it returns passes through whoever operates that server — and most people never read what that operator is allowed to do with it. A security scan tells you whether a tool’s description is hiding an instruction. It says nothing about whether the vendor retains your tool-call arguments for a year, feeds them into a training set, or hands them to a subprocessor you’ve never heard of. That’s a different document, and it’s worth five minutes before you click “Allow.”
Why this is a separate question from a security scan
A scan looks at the server’s tool descriptions — whether they contain hidden instructions, suspicious phrasing, or a mismatch between what a tool claims to do and what it’s actually allowed to touch. That tells you whether the server is trying to trick your AI client. It doesn’t tell you what the server’s operator does with the data once a legitimate tool call succeeds. Those are two separate trust questions: is this tool honest about what it does, and what happens to the data after it does it. The second one lives in the privacy policy, not the tool list.
Find the actual document, not the marketing page
Most vendors put their general privacy policy on a footer link, written for their whole product — account signup, billing, support tickets — and it may not mention MCP, tool calls, or AI access at all. Look specifically for a Data Processing Addendum (DPA), a subprocessors page, or a dedicated AI/API data section. If the MCP server is a side project or community wrapper rather than the vendor’s own product, there may be no policy at all — which is itself the answer to the question, just not a reassuring one.
Six things worth checking
1. Is tool-call data used for training? Some vendors train their own models on product usage by default, with an opt-out buried in account settings. If the server itself calls out to a third-party LLM to do its job, a second training policy applies — the one belonging to whichever model it uses, not just the vendor in front of you.
2. How long is data retained, and where? “We retain logs for security purposes” can mean thirty days or three years. Arguments to a tool call often contain the actual content you’re working with — a customer record, a draft contract, a support ticket — not just metadata about the call.
3. Who are the subprocessors? Hosting providers, analytics vendors, and the underlying model provider (if the server calls one) are all additional parties your data passes through. A subprocessors list, if the vendor publishes one, is usually more informative than the policy’s prose.
4. Does the policy even address the AI access path? A policy written before the vendor shipped an MCP server may only cover its existing product surface — the web app, the mobile app, the REST API. Tool calls from an AI client are a newer channel into the same backend, and an older policy may simply not have been updated to mention it.
5. Is there a DPA if you need one? If you’re moving regulated data — personal data under GDPR, health data under HIPAA, or anything inside a DORA-scoped financial entity — you need a signed data processing agreement with this vendor, the same way you’d need one for any other processor. See the GDPR and HIPAA posts for what that agreement actually needs to say.
6. What happens when you revoke access? Revoking an OAuth grant stops new tool calls. It rarely deletes what was already retained. If you need data actually removed after an engagement ends, that’s a separate request — and the policy should say whether the vendor honors one.
Red flags worth pausing on
- A server using a well-known brand’s name with no privacy policy, no company information, and no way to confirm who actually operates it — see how to tell a real server from a fake one.
- A local server run from an unfamiliar package with no stated operator at all, where the real question isn’t retention but what the process can already see on your machine — covered in MCP server supply chain risk.
- A policy that reads as if the AI access path doesn’t exist — no mention of API or tool-call data distinct from the product’s normal usage.
A five-minute version of this checklist
- Find the actual policy or DPA for the server, not just the vendor’s general product page.
- Check whether tool-call data can be used for training, and whether that’s opt-out or opt-in.
- Note the stated retention period and whether it applies to call arguments, not just metadata.
- Look for a subprocessors list, especially a third-party model provider if the server calls one.
- Confirm a DPA is available if you’re moving regulated data.
- Ask what happens to retained data after you revoke access — and get that in writing if it matters.
None of this replaces the usual vetting checklist for tools and permissions — it runs alongside it. A server can pass every security check and still be a bad place to send real customer data, simply because of what its operator is contractually allowed to do with it afterward.