Skip to content

Secure an AI's access to your API

Updated October 29, 2026 · Published October 24, 2026

To secure an AI's access to your API, expose only the tools it needs, prefer read-only ones, give it a dedicated credential with minimal rights, keep the MCP key secret and watch the quota and the call log.

1. Expose few tools, mostly read-only

Voozeo marks every tool by its HTTP method:

TypeHTTP methodsWhat the AI client is told
Read-onlyGET, HEAD, OPTIONSThe tool does not change data
WritePOST, PUT, PATCHThe tool changes data
DeleteDELETEThe tool is destructive

These hints let clients ask for confirmation before a risky call. Go further and disable, in the "Tools" tab, every write or delete tool that your use case does not need. An AI chooses better among ten clear tools than among fifty. See choosing and describing tools.

2. Give Voozeo a dedicated credential

In the "Target API" tab, Voozeo stores one secret for your API: an API key, a bearer token or a basic login. Create a credential just for this server, with the minimal rights, and never reuse an administrator credential. If it leaks, you can revoke it without touching anything else. If your API supports OAuth 2.0, choose it instead: each person then signs in with their own account and rights, and no shared credential is stored. See OAuth 2.0.

The secret is stored encrypted, can only be written and is never displayed again. The AI never receives it. See target API and authentication.

3. Protect the MCP key

The MCP key identifies the AI client to your server. It starts with vz_ and is shown only once.

  • Store it in a password manager, never in a repository or a message.
  • Prefer the URL with the key in the Authorization header (Claude Code, Cursor, VS Code). The URL that contains the key (claude.ai, Claude Desktop, ChatGPT) is a secret too.
  • Use a distinct server, hence a distinct key, for each use case.
  • If the key is lost or leaks, regenerate it: the old one stops working immediately.
  • On an OAuth 2.0 server there is no key: each AI signs in, and "Revoke all authorizations" in the "Target API" tab disconnects them all. See active authorizations.

A missing or wrong key gives a 401 error, and a server that is not "active" gives a 403 error. See the MCP key.

4. Rely on the built-in limits

SafeguardEffect
QuotaCalls above your plan's monthly limit are not executed, and your API is not called. The Free plan includes 1,000 calls per month.
Disabled serverThe "Server enabled" switch refuses AI calls at once, and keeps your sources, tools and key.
Private addresses refusedVoozeo refuses private and internal addresses, for a spec URL and for tool calls (UNSAFE_URL).
Truncated responsesA response above 100 KB is truncated before it reaches the AI.

The workspace owner is warned by email at 80% and at 100% of the quota.

5. Treat documents as untrusted

If your server also holds documents, remember that a document can contain instructions. Voozeo passes the text to the AI as data to consult, not as commands. Even so, do not upload documents that you do not trust, and do not expose write tools on a server that reads documents from outside sources.

6. Check what happens

Each call appears in the "Log" tab with its date, tool, status, HTTP code and duration. By default, the log never records arguments, results, your MCP key or your API secrets, so it is safe to review. The optional diagnostic mode, off by default, keeps the masked content of calls for 30 days; your MCP key and the request headers are never kept. The dashboard shows requests per day, the error rate and the AI clients that call your server. See call log.

Checklist

  1. Only the tools you need are enabled, and most are read-only.
  2. The target API credential is dedicated and has minimal rights.
  3. The MCP key is in a password manager, not in a repository.
  4. You know your plan's quota and when the alert emails are sent.
  5. You review the log after the first week.

Go further

Get started