OpenAI's Responses API documentation says that with store set to true, data is retained for 30 days unless zero data retention is enabled, and that MCP servers are third party services governed by their own retention and residency policies. For security data, that means two retention regimes to check: the model provider's and the connector vendor's.
What the documentation states
The guide for remote MCP servers and connectors says three relevant things. First, with store set to true, data is retained for 30 days unless Zero Data Retention is enabled. Second, MCP servers are third party services with their own data retention and residency policies. Third, the API does not store the authorization value and does not show it in the response object, so you send it with each request.
The page does not describe your agreement, your plan's settings or any other product's behavior. Retention terms differ by product and plan, so check the terms that apply to you.
Two regimes, two questions
| Regime | Question to ask |
|---|---|
| Model provider | How long are requests and tool results kept, and can I turn on zero data retention? |
| Connector vendor | What does the server log, how long, and where? |
Tool results include whatever the connector returns, such as project names, package names and findings. If you store responses, those results sit in the provider's storage for the stated period. If the connector vendor logs the arguments, those sit in the vendor's logs.
A worked example
This example is illustrative. A team builds an internal assistant over a vulnerability platform using the Responses API. They go through the checklist.
- They list the fields the tools return and classify them. Package names are low sensitivity. Internal project names and owner names are higher.
- They check their provider agreement for zero data retention, and note the setting for storing responses in their request code.
- They ask the platform vendor for log retention and location, in writing.
- They decide not to store responses for the pilot, and record the decision.
- They set the approvals option so a person approves any tool that sends data out.
- They add a line to the privacy notice for internal users saying what is processed and by whom.
Step by step: a retention review
- Identify every system that sees tool results: model provider, connector vendor, your own logs.
- For each, record retention, location, access and deletion route.
- Decide whether any data class is excluded from the assistant.
- Check whether your request settings store responses, and turn that off if you do not need it.
- Confirm that tokens are never logged by your own code.
- Test deletion: ask each party how to remove a specific conversation.
- Review annually or when terms change.
Common mistakes
Teams often review only the provider and forget the connector's logs. They also leave response storage on because it is a default or because it helps debugging, then forget it exists. Another mistake is putting sensitive text in tool definitions, which the documentation advises against.
Do not promise what you cannot prove
When you answer a customer questionnaire about AI data handling, state what you configured and how you checked it, and cite the documentation date. Do not state that data is never retained unless you have the settings and agreements to back it.
Where Safeguard fits
Safeguard offers compliance and evidence automation and an MCP Guardrail for teams that need to show how agents are controlled. The free scan is a no assistant way to start. The Safeguard CLI, installed with a curl or PowerShell installer, can also write code, and the download page lists a Local agent and a Cloud agent as available.
Wording for your documentation
Write the answer once and reuse it. A short paragraph might say: our assistant uses a model provider under an agreement we reviewed on a stated date. Response storage is set to the value recorded in our configuration. Tool results come from a connector operated by a named vendor, whose retention terms we obtained in writing. Users are told not to paste secrets or customer data. Update the paragraph whenever a setting or an agreement changes.
Questions to put to the connector vendor
Ask what the server logs per call: user, tool, arguments, result size, or the result itself. Ask how long logs are kept, where, and who inside the vendor can read them. Ask whether arguments or results are used for anything beyond serving the request, and whether you can request deletion of a specific period of logs. Ask for the answers in a document, not on a call, and file it with the connector register.
If the vendor logs full results, treat the vendor as holding a second copy of whatever the assistant retrieved, and apply the same classification you would to any other copy of that data.
Try it yourself
Click through it at your own pace, or open it full screen.
Trace which steps in the conversation would send data to a model provider and which to a connector.
Quick check
Does zero data retention apply to every setup?
Not automatically. The page says retention is 30 days with store set to true unless zero data retention is enabled, so check your own settings and agreement.
Who keeps the data the connector logs?
The connector vendor, under its own policies. The documentation says MCP servers are governed by their own retention and residency policies.
Does the API store my access token?
The documentation says it does not store the authorization value, so you send it with every request.
Sources
- OpenAI documentation, remote MCP servers and connectors in the Responses API
- Safeguard download page
Last reviewed: 19 October 2026.
