Two related features, worth keeping distinct.

Copilot

A conversation scoped to one site, with the site's own live data in context — its versions, updates, alerts, metrics and recent history. Because it is looking at the real site, questions like why is this one slower than the rest or what changed before the errors started get grounded answers rather than generic advice.

Copilot reads. It does not change the site by itself.

AI tasks

A job you hand over: describe what you want checked or done, and it works against the site and reports back. Tasks that would change the site require your confirmation before they act.

Credits

Both draw on AI credits, which are consumed per use and shown in the sidebar. A task that fails without producing an answer does not consume credits.

What to expect

These are assistants working from real data, which makes them good at what is going on here and unreliable as the last word on anything that matters. Treat a recommendation as a colleague's suggestion: worth reading, worth checking.