Security

Credential handling

How an agent's secrets are stored, the one path that stores a secret without the AI model reading it, and how platform keys are kept.

Agent secrets

A credential your agent needs, such as an API key for a booking system, is stored as a secret variable.

StoreWhat it holds
The variable recordThe secret's encrypted value
A separate key serviceThe key that decrypts it, held in Sydney

Access

Only our servers can decrypt a secret, and only for the workspace and record it was stored under. No browser, API token or signed-in user can read a value back. Someone without access to the agent sees no sign that the secret exists.

Deleting a variable deletes its value from that version. Earlier versions keep their own copy until they are deleted. Encrypted secret values are included in our database backups, which are kept for seven days in Sydney.

How a secret reaches storage

One path keeps a credential out of the AI model: the dialog. The in-app assistant asks for a secret by name, you type the value into a dialog, and the dialog sends it to our server, which encrypts and stores it. The value never enters the assistant's context.

The assistant gets back a confirmation carrying the variable's name and nothing else. No tool on that path takes the value itself.

The workspace assistant cannot do it

The workspace assistant works across agents rather than one open agent. It does not render the dialog, so it cannot store a secret. It can still write plain fields, so it is not a safe place to paste a credential either.

What the dialog protects you from

The chat itself. Anything typed into a conversation is read by the model before any tool runs. Several agent fields also store values exactly as written:

  • The default value of a variable.
  • A tool's request headers.
  • A tool's endpoint URL.

Ask for a credential to be put in one of those and it can end up stored in the clear. The assistant refuses some such requests and names the placeholder to send instead, but the refusals recognise a credential by its shape, and a credential does not always have one. The dialog is the only path we guarantee: the value goes from your keyboard to encrypted storage without passing through the model.

If you have already pasted a credential into a chat

Rotate it, and check the agent's variables and its tools for a copy left behind.

The MCP server and the operations API

Both accept a secret value directly. The caller supplies it deliberately and there is no dialog to render. The value transits whatever client sent it.

Platform keys

API keys and publishable keys are stored as hashes. The raw value is shown once at creation and cannot be retrieved afterwards.

On this page