FAQ

Frequently asked questions

If yours is missing, write to us: we answer, and we add the question.

Does masking degrade answer quality?

This is exactly the difference between substituting and deleting. A tool that blacks out a value leaves a hole: the model loses the thread, and so does its answer. Neverseen writes a value of the same nature and shape in its place, consistent throughout the prompt. The model reasons over a complete story — it simply is not yours.

How much latency does the agent add?

Detection and substitution run locally, over the text of the request, and are measured in milliseconds. The dominant cost stays the provider’s own round trip, several orders of magnitude above.

What does your backend see?

Counters and configuration state: request count, masked values per category, token volumes per model, requests per provider and per AI tool, tool calls reduced to a closed vocabulary, conversations described as histograms, and what the agent actually applies — version, locales, mode, switched-off categories, secret level, and how the agent itself is exposed. Never a prompt, never a response, never a detected value, no file name and no URL. Two fields are personal data and are named as such: the machine’s own local IP addresses and the name it calls itself. The telemetry contract is public and a test forbids any free-text field that has not been argued for.

What happens if your service goes down?

Nothing, from the user’s point of view. The agent does not wait for the backend to mask: monitoring is an observer, not a checkpoint. Heartbeats accumulate locally and leave when the link is back.

Is it open source?

It is source available. FSL-1.1-Apache-2.0 makes the code public and its use free and unlimited, commercial use included, while forbidding competing products. Every release becomes Apache-2.0 after two years. We do not claim the OSI open source label, because it does not apply.

Are secrets handled like everything else?

Almost. A secret is detected and masked like any category, and it is restored on the way back like any other value — otherwise your tool would receive a token where it expects a key. Two rules do set it apart. It never gets a stand-in: it always travels as a token, even in stand-in mode, so no response can carry a plausible-looking key nobody ever wrote. And it cannot be switched out of the policy, under any name: a secret category somebody could disable would be a key sent to a provider in clear.

Which tools work with the agent?

Anything that talks to a provider over HTTP and accepts a configurable base URL: Claude Code, AI CLIs and SDKs, your internal scripts, the official libraries of the eight supported providers.

Can it be rolled out fleet-wide?

Yes, through your usual endpoint management tooling: the install script is non-interactive and enrolment uses a token you distribute. The console then shows you the machines that never reported.

What about data already sent before installation?

Neverseen does not recover it. No tool can. What it gives you is a date from which the control is in place, and continuous proof that it has run since.

How is this different from an enterprise gateway?

A gateway protects the applications you built and pointed at it. Neverseen protects endpoints, and therefore how your people use AI — the gateway’s blind spot. The two approaches complement each other more than they replace each other.