Security

The REST API and the MCP server send no CORS headers, so a browser page cannot call them, and the MCP server refuses any request that comes from a browser page. Keep keys on a server, never in a browser, a mobile app or a public repository. Read them from an environment variable or your secret manager, as every sample here does. Never write the Authorization header or a token to a log, and keep an app's OAuth tokens on its server too, never in the browser's storage.

• For a key: its first few characters and a one-way fingerprint (a SHA-256 hash). Never the key itself, so nobody, Current included, can show it again.
• OAuth access and refresh tokens and app client secrets: the same, a fingerprint only, each kind in its own space so one can never be used as another.
• A webhook signing secret is the one secret Current must read back, to sign each message. It is stored encrypted and shown to you once.
• The call log keeps which key or app called, what it did, its result and its address. Never a request body, a reply's words or a key.

• Typed prefixes and a checksum, so a secret scanner can recognise a Current key and a typo is refused before any lookup.
• Every key expires, and one unused for 30 days switches itself off.
• A key that changes tickets works only from its allowed addresses.
• Every refused key gets the same 401, so a guess learns nothing.
• Alerts on the API access page: a key or an app used from a new address, a sudden run of refused calls, or a burst of changes.
• Your tenant admins see every live key in the workspace, people's personal MCP keys included, and can revoke any of them. They never see a key itself.

1. 1
   Revoke it

   On the API access page, revoke the key. It stops at once.
2. 2
   Check the call log

   Filter the call log by that key to see what it did and from which addresses.
3. 3
   Make a new one

   Create a new key for the agent, with an allowlist, and store it in your secret manager.

Current delivers webhooks only to public https addresses on the standard port, checks the address on every try and never follows a redirect, so a webhook cannot be pointed at your own network. Text from people reaches an AI assistant labelled as untrusted data (see MCP and AI assistants).