-
Notifications
You must be signed in to change notification settings - Fork 0
Reference Privacy
Agent Karma Docs edited this page Jun 13, 2026
·
2 revisions
📖 This page mirrors the canonical Agent Karma docs. Edits should be made there, not in the wiki.
The Agent Karma trust contract — what's captured, what never is, and your controls.
Agent Karma is local-first by design. Seven promises, hardcoded into how it works:
Local-first. No source code captured. No terminal output captured.
No cloud upload. No telemetry. No login. No surveillance.
-
File-change events — file names only (including edits made by AI agents
or the CLI, gated by
agentKarma.captureExternalFileChanges). Never file contents. - Validation command types — that you ran a Test / Build / Lint / Type Check. The raw command string is classified, then discarded.
- A git diff summary — file and line counts, never the diff content.
- The short SHA of commits — never the message, diff, author, or files.
-
Your typed intent text — recorded locally by default (so your cards and
dashboard read back meaningfully). Turn it off with
agentKarma.capturePromptText; either way it never leaves your machine, and the prompt-clarity score never reveals it.
- Your source code.
- Your terminal output.
- Your keystrokes.
Local JSON under VS Code's per-extension storage, on your machine. No backend, no account, nothing leaves your computer. The extension makes no network calls — a Prime Directive enforced in CI.
- Export the current session as JSON or Markdown (metadata only).
- Reset Karma History — clears sessions and the trend, keeps your settings.
- Delete All Local Data — wipes everything, one command.
- Per-setting granularity over what's captured at all.
Uninstalling removes the tool; for a privacy-first coach, your local data going with it is the point.
Start here
Concepts
Features
- Everything it does
- Sessions & the Dharma Card
- Your cards: Dharma, Karma & Phal
- Dashboard & cards
- Readiness, nudge & chat
Reference
Guides