Terminal
Terminal/Skill
Checks whether a live event fires in Amplitude, measures the impact of a product issue, and finds the affected users for incident analysis.
By Amplitude
Works with
npx skills add amplitude/mcp-marketplace --skill live-data-forensicsThe most common multi-tool job on this MCP server: verify live event behavior, quantify impact, find affected users. Calls observed in the wild follow one arc — follow it too.
get_amplitude_context (no args) → again with projectId.
Do not guess project IDs.search_amp_entities for existing analyses of the same
area — saved charts already encode correct event names and segments.check_for_recent_event_ingestion confirms
first-seen/last-seen before you query — a silent event means the chart
will be empty no matter how correct the definition is.query_amplitude_data bursts — one slice per call (by
version, by reason property, by platform), not one mega-query. Prefer the
typed chart parameter (kind: 'segmentation', events + where/group_by +
date_range); it compiles server-side and validation + taxonomy checks run
automatically — no separate pre-flight call needed. Compare prod vs
staging/UAT projects when the question is environment-specific.query_amplitude_data with a user-ID group_by
to rank affected users → use_amplitude_cohorts action: 'find' for the
full set.get_amp_user_data include: 'timeline' per
user, batched (10–20 parallel calls is normal for population analysis;
the tool accepts up to 10 identifiers per call).interval ("Last 3 Years" at weekly interval — re-denominate as
"Last 156 Weeks" or change the interval); funnel conversion_window
missing unit; unknown filter operator for that chart kind (use set
for presence — works in every kind).date_range explicitly, either
{relative: "Last 30 Days"} or {start, end} epoch seconds, never both.
Sub-daily intervals only allow short windows (hour caps ~8 days); daily
granularity caps around 30 days.op
and scope from the taxonomy lookup, not intuition. A scope mismatch
("property X is not tracked on this event_type") means you used a user
property as an event property or vice versa.where (property
conditions) plus performed ("users who did event ≥N times in a window").
Omit segments entirely for all users.definition fallback (composition, revenueLtv, advanced params):
on failure the response embeds the chart-type schema with valid enums and
a working example — fix from that and retry.read ETIMEDOUT is a backend timeout — narrow the date range/filters and
retry once.