Anthropic-Claude Code-onboarding — Claude Code skill

What xai-grok-onboarding does

The skill walks through a standard OAuth 2.0 flow (RFC 8628 Device Authorization Grant). No API key is involved: model access comes from the subscription itself. It runs in script mode and registers no tools, so you read SKILL.md and call the exported functions from a bash block.

Four steps are visible to the user. start() hands back a verification_url_with_code; the user opens it, signs in and approves. Then poll() reports one of three outcomes — connected, pending, or a failure. Polling early is a waste, since a pending state still costs tokens. Once connected, the user has to reload the page by hand: the model list is cached client-side and does not refresh on its own. Models show up under the xai-grok/ prefix, with xai-grok/grok-4.3 as the default.

One catch sits on Anthropic's side. Some accounts get access_denied from the token endpoint even with an active SuperGrok subscription, and the gate is deterministic per account — retrying gets you nowhere. The fallback is the byok-custom-model skill with a key from console.x.ai. The skill also leans on the core.xai_grok package inside the platform image; on older images it loads but every function answers xai_oauth_backend_unavailable, which means the image needs updating rather than another attempt.