These are not aspirations. Each one is either enforced by the code or checkable by you, and Privacy tells you how to check.
Nothing you say leaves the machine. No account, no login, no analytics, no crash reporting, no "anonymous usage statistics" sent anywhere. The only outbound request Kalamos ever makes is the one-time model download on first run. After that it works with the network off — try it.
What touches the disk, precisely. Two things, both local, both inspectable, because "we store nothing" is usually a lie and you should not have to take it on faith:
- The last 25 transcriptions, in the app's own preferences
(
~/Library/Preferences/com.kalamos.app.plist). They exist so a mis-fired paste or a wrong focused field never loses what you said: they are listed in the menu, one click copies any of them back, and Clear History deletes them all. Never transmitted, never more than 25. - A timestamp-only usage log — the time you dictated, never the text — so
Scripts/analyze-idle.tscan suggest an idle timeout that fits how you actually work.
Full transcript logging exists for debugging, is off by default, and --doctor
keeps warning you for as long as it is on.
No subscription, no tiers, no upsell. MIT licensed. There is no paid version to be nudged toward, because there is no paid version.
It fails loudly or not at all. A silent fallback that returns plausible text
is worse than an error, because you cannot see it. This is not theoretical: a
substring bug once classified iTerm2 as a code editor — its bundle id is
com.googlecode.iterm2, and "googlecode" contains "code" — so AI cleanup was
skipped in the terminal for months, with no visible symptom at all. That class of
failure now has tests, and --doctor exists to make the rest of it visible.
Your Mac is not a datacenter. Both models unload themselves from memory after you stop dictating and come back in a few seconds. The idle timeout is yours to set, including "never".