Skip to main content
The CLI assumes the caller might not be a human. That shows up in how it returns data, where it keeps its state, and how it signs you in.

​
The JSON envelope

Pass --json as a top-level flag, before the subcommand, and every command returns a structured envelope:
success: true always carries data. success: false always carries error. List endpoints ship pagination alongside the data: pagination: { hasMore, nextCursor }.

​
Context discovery

The CLI keeps its state in two files an agent can check before acting:
If .gobi/settings.yaml is missing, gobi vault init and gobi space warp are the interactive entry points. They need user input, so an agent should hand off to the user instead of driving them silently. Any gobi space command also takes --space-slug <slug> to override the default without touching settings.

​
Headless auth

gobi auth login prints a URL and a user code to stdout, then polls until the user authorizes. An agent can run it as a background task, surface the URL as a clickable link, and wait for the process to exit.

​
Errors worth planning for

With --json, every failure is success: false plus a human-readable error. Three come up in normal agent use:
  • Not signed in. Check gobi --json auth status before a run; login is interactive, so hand off to the user.
  • Owner-only content. transcript and audio belong to the person who recorded. Another member’s returns a 403 by design. Don’t retry; there is no flag around it.
  • Vault deletion refused. gobi vault delete is rejected while the vault still owns posts, members, or files.

​
Claude Code skills

The package ships a .claude-plugin/ manifest with skills wrapping each command group: