MCP key safety checklist for Safeguard workspaces

An MCP key acts as the member who created it, across the whole workspace, so treat it like that person's password. Create one only when OAuth is impossible, store it in a secrets manager, keep it out of source control, give each client its own key, and revoke it the moment it is no longer needed.

What the key is

In Safeguard, you open Settings, then Developer and create a key of the type MCP key. The client sends it as an Authorization header with the value Bearer followed by an sg_sk_ prefixed string. The key is shown once, so you cannot look it up later.

The key does not bring new powers. It lets a client do what its creator can do, limited by the tools the workspace has enabled. But a client holding the key can do it without any further sign in.

The checklist

1. Prefer OAuth. If your client supports it, skip the key. Claude, ChatGPT, Cursor, VS Code and Gemini CLI all do. See OAuth or MCP key.
2. One key per client. If two tools share a key, you cannot revoke one without breaking the other, and you cannot tell which one misbehaved.
3. Create it as the right person. The key inherits its creator's role. A key made by an administrator carries administrator reach. Where possible, create keys from an account with only the access the client needs.
4. Keep it out of source control. In Cursor or VS Code, put it in your user level configuration, not a project file you commit. Add the file to your ignore list if there is any doubt.
5. Use a secrets manager for anything run by a script or pipeline.
6. Keep Custom GPTs private. A Custom GPT Action uses one pasted key for every user of that GPT.
7. Name and record it. Note who owns each key, what client uses it, and when you reviewed it.
8. Review on a schedule. Remove keys whose client you no longer run.
9. Revoke fast. Revoke under Settings, then Developer. It takes effect within a minute.
10. Offboard people. A key whose owner's account is closed stops working within a minute, but do not rely on that alone. Revoke it in the offboarding checklist.

What stops a key automatically

The server checks each key at least once a minute. A key stops working if it is revoked, if its owner's account is closed, or if the workspace turns off API-key access to MCP. That last switch is a useful emergency lever: it blocks every key at once.

What to do if a key leaks

1. Revoke it under Settings, then Developer.
2. Create a replacement only if the client still needs one, and store it properly.
3. Review recent activity by that member for anything unexpected.
4. Find how it leaked, such as a committed file or a shared GPT, and fix that path.

Do not install look-alikes

Safeguard publishes no npm package or local bridge for its MCP server. The server never reads SAFEGUARD_API_KEY from your machine. A package that asks you for a key and claims to be the Safeguard server is untrusted.

Quick check

Sources

• Safeguard docs: MCP integrations overview, authentication
• Safeguard docs: Cursor guide
• Safeguard docs: ChatGPT guide

Last reviewed: 19 October 2026.