Skip to main content
SelfTune starts local. Installing the skill, reading local agent histories, installing hooks, running a background service, and connecting Sync & Backup are separate actions with separate permission boundaries.

​
Local data

Local SelfTune can read supported agent histories, discover skill packages, and store normalized evidence in its local state directory. The local dashboard and CLI read the same data. A local quickstart does not require remote upload. Treat agent histories as sensitive even when SelfTune normalizes them. Limit filesystem access to the histories and skills required for the chosen workflow.

​
Optional remote data

SelfTune Cloud and self-hosted Sync & Backup connections are opt-in. Before connecting, review the destination, account, artifact types, retention, and recipient scope. Raw transcripts are never synced; skill revisions, selected drafts, Skill Sets, decisions, and redacted evidence summaries may be.

​
Actions that require explicit approval

  • installing hooks or importing historical sessions;
  • enabling a persistent service or recurring schedule;
  • creating an account, token, or credential;
  • uploading evidence or skill artifacts;
  • publishing or sharing a skill;
  • enabling paid services or automatic mutations; and
  • deleting local or remote data.
Receiving a shared skill can involve several independent actions:
  • opening a claim link previews metadata and does not install anything;
  • claiming or importing adds an immutable package to a SelfTune library and does not change the local filesystem;
  • downloading returns a portable package; and
  • applying an imported Skill Set previews the affected agents, projects, paths, and conflicts, then waits for confirmation.
License acceptance does not imply telemetry consent. Optional lifecycle reporting tells the sender only that a disclosed action occurred. Optional contributor signals are default-off and use a separate version-scoped routing capability. Raw prompts, transcripts, files, code, local paths, email, machine identity, and stable user identity are not included. See Skill distribution, licensing, and signals for the complete model.

​
Before approving a remote action

Ask the agent to show the exact destination, identity, files or fields, and retention behavior. Prefer a preview or manifest. Keep raw prompts, transcripts, credentials, and unrelated project files out of skill packages.