Mask an n8n workflow's texts

The Latet nodes for n8n (@latet-ai/n8n-nodes-latet, under the MIT license) call the applications API of a Latet server from a workflow. They scan, mask, unmask and forget the texts of the workflow’s items: a workflow masks what it sends to a model or to a service, and unmasks what comes back. They need a Latet server, 0.10.0 or later, and an API key’s token; a person’s computer token is refused.

Install the nodes

In n8n: Settings > Community Nodes > Install, then enter @latet-ai/n8n-nodes-latet. n8n’s installation guide has the details.

A token for the workflow

The administrator of the Latet server makes a token with the permissions the workflow uses and no more (API keys and tokens):

latet server --server https://latet.acme.lan keys add-token "Support desk" n8n --mask --unmask --out n8n.token

Operation	Permission
Scan	scan, or mask
Mask	mask
Unmask	unmask
Forget	forget, which no other permission brings

In n8n, the Latet API credentials hold the server’s address (https://latet.acme.lan) and the token, the content of n8n.token. Use an https:// address: over http:// the token and the texts travel in clear. n8n tests the credentials with a scan of one word, which keeps nothing and needs scan or mask: a token that may only unmask or forget fails the test and still works in the node.

What each operation does

Each operation reads the text of each item from a field (Text Field, text by default) and sends the items in calls of at most 1,000 texts.

• Scan says where the values are in each text, by type, and keeps nothing.
• Mask gives each text with fakes in place of its values, the same fake for the same value in every call of the key. Kind says what the texts are: Prose for text people wrote, Data for logs, records and a program’s output (prose and data).
• Unmask puts the real values back in place of the key’s fakes. The server writes each unmask in its administration log.
• Forget takes values out of the key’s vault, one per item (Value Field, value by default), written as they were masked (forgetting values).

A text the server does not answer, because its local reviewer stopped it or its guard against dumps withheld it, comes out with blocked and no text, so that the text in clear goes no further.

A model that never sees the values

A workflow that answers support messages with a model:

1. A trigger receives the message.
2. Latet: Mask, kind Prose, masks the message.
3. The model’s node (OpenAI, Anthropic, Gemini, an AI Agent) gets the masked message in its prompt.
4. Latet: Unmask, with Text Field set to the field that holds the model’s answer, puts the real values back.
5. The workflow sends the answer.

The model sees fakes only, the same ones across the messages of one customer, and the customer reads the real values. Do not give an AI agent the Unmask operation as a tool: it would hand the model the very values masking keeps from it.

The package’s README lists every field the nodes write and the errors they give.