On-device classification

Deep inspection.
On the device.

To know whether a prompt carries source code or customer PII, something has to read it. Northbeams reads it on the laptop, labels the risk, and sends out only the labels. That is how you get depth without a proxy, and privacy without giving up detail.

See your shadow AI free →Book a demo

Browser · Desktop · CLI · MCP · Network

The mechanism

The classifier runs where the prompt already is.

A gateway has to route your traffic and break TLS to read a prompt. Northbeams does not. The prompt is already on the device, so the classifier reads it there, decides the category and severity, and only the verdict leaves. No traffic to route, no certificate to install.

The line, drawn plainly

What leaves the device, and what never does.

This is the whole privacy model in two lists. The left is what the console receives. The right is what stays on the laptop, always.

WHAT LEAVES THE DEVICE

Category labels (credentials, PII, source code, customer data, contracts). Severity. Event counts. A short redacted, hashed snippet for context. Per-user attribution: which tool, which user, which time.

WHAT NEVER DOES

Raw prompts. MCP argument values. Keystrokes. Mouse movement or idle time. Non-AI browsing. Slack, email, or calendar content. And there is no productivity scoring of any kind.

Why it is worth the effort

Depth a hostname cannot give you.

It reads the risk, not the domain

A network tool can tell you someone reached a chatbot domain. On-device classification tells you the prompt carried a customer list, flagged high, and names the user. That is the difference between a log line and a decision.

Flagged with severity and value

Each finding carries a category, a severity, and an estimated exposure value, so you can triage by what actually matters instead of scrolling through raw text.

No proxy, no MITM cert

Because inspection happens on the device, there is no traffic to intercept and no certificate to install. Nothing in the network path to break, and nothing to slow the user down.

You hold the dial

Privacy modes the admin controls.

The default is conservative. Where you need to go tighter or looser, the controls are yours, set centrally and applied per team.

→Turn the hashed snippet off entirely, so only labels and counts leave
→Set how aggressively values are redacted before anything is stored
→Scope inspection by team, so rules match how each group actually works
→Choose the retention window: unlimited history, or a fixed window on Enterprise

Reversible and auditable

Every policy change and every response action is logged to one signed, immutable record. What you turned on, when, and by whom.

The hard questions

Asked directly

Does the raw prompt ever leave the laptop?

No. The classifier runs on the device and the raw prompt stays there. What leaves is the category, the severity, counts, and an optional short hashed snippet.

What is in the hashed snippet?

A short, redacted fragment for context, hashed so it is not a readable copy of the original. You can turn it off entirely and send only labels and counts.

Is this employee monitoring?

No. Northbeams does not capture keystrokes, mouse or idle time, non-AI browsing, or the content of Slack, email, or calendar, and it does no productivity scoring. It watches AI use, not people.

Does it read MCP argument values?

It classifies them on the device to decide allow, warn, or block per tool. The argument values themselves do not leave the laptop.

Depth without the man in the middle.

See how on-device classification maps your real risk. Free to see, no network change to start.