How RTK works

Before reading the dashboards, it helps to know what actually happens on a developer's machine and how a number ends up on your screen.

The idea in one sentence

AI coding assistants read the output of the commands they run — git status, test logs, file dumps — and that output is often huge and mostly noise (ANSI colors, progress bars, repeated boilerplate). RTK filters that output on your machine before it reaches the AI, so the model receives a compact version. Fewer tokens in means lower cost and more room in the context window — without changing what you do.

What "tokens saved" means

Every command produces a raw amount of output and a compact amount actually sent to the AI. Tokens saved = raw − compact: the noise RTK removed before it reached the model. It's measured on your machine by rtk and reported as a single number per command — the dashboards then sum and price it.

Secret values never leave your machine

If a command's output contains a secret (an API key, a token…), Shield redacts it locally, before anything is sent to the AI or reported to the cloud. The secret value itself never leaves your machine. So the dashboard can group and rank leaks, RTK does report non-reversible details about each detection — a one-way fingerprint, a masked preview (just the type/prefix, like sk_***), and where it happened (command, repo) — but never the secret.

What the cloud stores

Just the telemetry: per command, the token counts, the command name, the AI tool, the project/session labels, and who ran it (your account). That's enough to build every chart in Analytics and Savings — and nothing sensitive.