Connect Claude to Amplitude

List the event types visible in the project, each with a weekly total. Call this first to learn the exact event names the segmentation, funnel and retention actions need - Amplitude matches event_type literally, and a near-miss returns an empty series rather than an error. Hidden events are excluded.

The general-purpose reporting action: count an event over a date range, optionally against a second event, segmented and grouped. `metric` picks what is counted - uniques (distinct users), totals (all occurrences), pct_dau, average, or formula for a ratio such as UNIQUES(A)/UNIQUES(B), which needs the formula argument and a second event. Dates are YYYYMMDD, both required. `interval` is Amplitude's own coding: 1 daily, 7 weekly, 30 monthly, -3600000 hourly, -300000 real-time, and the shorter the interval the shorter the window Amplitude will serve - real-time reaches back two days, hourly seven, daily a year.

Run a conversion funnel over an ad-hoc list of steps. Give the steps in order as step_1 through step_5; only the first two are required and the rest are sent only when supplied. `mode` decides how strictly the order is read: ordered means the steps happened in that order, sequential means with nothing else in between, unordered means in any order. `conversion_window_seconds` is how long a user has to finish, defaulting to 30 days. Dates are YYYYMMDD.

Measure how many users who did a starting action came back to do a returning action. start_event takes the literal _new for new users or _active for all users, as well as a real event type; returning_event takes _all for any event or _active for any active event. `retention_type` defaults to n-day; choose bracket to group days together, which then requires brackets, e.g. [[0,5]] for a Day 0 to Day 4 window. Dates are YYYYMMDD. Note the interval here has no hourly or real-time value - only 1, 7 or 30.

Count active or new users per day, week or month over a date range, optionally segmented and grouped. Use this for the plain "how many users" question, where event_segmentation would need an event to count. Dates are YYYYMMDD.

How many users are active right now, in five-minute buckets over the last two days, with yesterday alongside today for comparison. Takes no arguments - Amplitude fixes the window and the granularity.

Fetch one user's profile and their event stream, newest first. This returns a real person: their user ID, device IDs, country, region, city, last known coordinates and every user property on their profile, followed by the events they fired. Use it to explain one specific user's path, not to gather users - Amplitude rate-limits this endpoint and its sibling get_user_search separately from the rest, at ten concurrent queries and 360 an hour.

Find a user's Amplitude ID from an Amplitude ID, a device ID, a user ID or a user ID prefix. This is how you turn an identifier you already have into the amplitude_id that get_user_activity needs. Note that no match is not an error: Amplitude answers 200 with type "nomatch" and an empty matches array.

List the chart annotations on the project - the dated notes a team leaves on a release, an outage or a campaign launch. Read these before explaining a jump or a drop in any of the reporting actions above; the explanation is often already written down here. Filter by chart_id or by category, but not by both, and by a date window. Dates here are ISO 8601 with an offset, not the YYYYMMDD the query endpoints use.