Not on every request. The recommended pattern is to pull what the bot needs from Mixpanel into WordPress on a schedule, using Action Scheduler or a sync plugin, then point SleekAI at those rows. That keeps replies fast, keeps Mixpanel rate limits intact, and means the bot can still answer when the Mixpanel UI is briefly unavailable.

Anywhere SleekAI can read it. A common pattern is a custom post type called Events, with ACF fields for required properties, owner, and description. Another is a single JSON document stored in an options field. Both work; the wizard maps either into named variables the system message references at request time.

Yes, if you expose them. SleekAI reads data attributes and any postmeta tied to the current URL, so a Bricks or block-based template that publishes its event inventory into a meta field can answer questions about what fires on that page. The bot does not intercept Mixpanel network calls directly.

Only what you map. SleekAI exposes a defined set of named variables in the system prompt, not the entire database. For privacy-sensitive properties, omit them from the mapping or run a separate public bot that can only see a redacted schema. Internal admins can have a richer bot via display conditions on user role.

Yes. SleekAI does not care whether events fire client-side or server-side; it only cares about what your WordPress site stores about them. If your server-side pipeline writes event metadata back into a WordPress post type or option, the bot reads it from there. The integration is about documentation and Q&A, not about replacing the tracker itself.

Yes, if cohort definitions are stored in WordPress. Most teams paste a one-paragraph definition of each major cohort into a doc post type alongside the membership rule. SleekAI reads that and answers in plain English. For numeric cohort sizes, schedule a small Mixpanel export into a custom field and the bot can quote the latest snapshot.

Any model supported by the BYO-key providers SleekAI ships with: OpenAI, Anthropic, Google, and OpenRouter. Tracking plan Q&A is mostly retrieval over a small schema, so a fast cheap model handles it well. Reserve stronger models for ambiguous questions about funnel definitions or for explaining cohort logic to non-analysts.

Yes. Every chat is stored inside WordPress with the model name, token usage, and page URL. For analytics teams, this is useful in two ways. It surfaces repeated questions that signal a missing doc, and it gives a paper trail when a privacy review asks who got told what about event collection.