Claude Code Onboarding Skill — Connect Claude Code to CData Connect AI

This skill walks an existing Claude Code user through everything needed to start using CData Connect AI — a managed MCP server that gives Claude live, contextual, governed SQL access to hundreds of business data sources (Salesforce, HubSpot, Jira, NetSuite, SAP, Snowflake, Google Analytics, and more).

The user is assumed to already have Claude Code installed and signed in. Do not walk them through installing the CLI, creating a Claude account, or running npm install -g @anthropic-ai/claude-code. Start at Connect AI sign-up.

Auth model: Connect AI's MCP endpoint accepts two auth methods. The default path is a Personal Access Token (PAT) + Basic auth: the user generates a PAT in the Connect AI console and Claude Code sends it in an Authorization: Basic header. Crucially, the PAT never enters the AI session — Claude hands the user a registration command that they run in their own terminal, so the credential stays local and out of the conversation transcript. This works on every surface — interactive or headless — and does not depend on the OAuth authorization server whitelisting Claude Code's dynamically-chosen localhost callback (a redirect_uri / "Callback URL mismatch" there is a common cause of OAuth failures). Browser-based OAuth is offered as an alternative (further down in Step 2) for users who prefer short-lived, self-refreshing tokens and whose Connect AI OAuth app has the Claude Code callback URL whitelisted — Connect AI's MCP endpoint is a fully OAuth-compliant resource (advertises WWW-Authenticate: Bearer + protected-resource metadata, supports dynamic client registration, PKCE, and refresh tokens).

Flow order: Authentication identifies the user, not a specific dataset, so Claude Code can be wired up immediately after sign-up — before any data source exists. The order is: sign up → connect Claude Code to Connect AI → connect a data source → verify. (Authenticating with zero data sources is fine; getCatalogs simply returns empty until a source is added.)

Delivery style

Walk the user through the steps one at a time — never dump them all at once. After each step, end with exactly one line:

Reply next when done — or tell me what went wrong.

Keep it identical every step so the user only has to remember one word. Do not ask multi-part questions, do not list optional details to report, do not offer alternative reply phrasings — that defeats the purpose. The exceptions are Step 0 (state detection, a multi-choice question), Step 1 (where it can help to ask which sign-up method the user picked), and the PAT step in Step 2 (where the user runs a command in their own terminal and reports the result, rather than simply replying next).

Do not give a time estimate up front (e.g. "this takes ~10–15 minutes"). Time guesses can put users off before they start, and the actual time varies wildly depending on which data source they connect.