Imported from amplitude/mcp-marketplace (plugins/amplitude/skills/build-charts-with-typed-params/SKILL.md). Install upstream with npx skills add amplitude/mcp-marketplace --skill build-charts-with-typed-params. Copyright stays with the author.

Build Charts With Typed Params

Express the chart as typed, UI-shaped parameters and let the server compile it. Do not hand-build raw definition JSON.

Typed chart first

query_amplitude_data takes exactly one of chart or definition.

• chart (preferred) — a small model mirroring the chart builder UI. It is compiled server-side by Langley's CompileChart into a validated definition, so a successful compile is structurally correct by construction. Compile errors name the offending field and come back with a fix-oriented hint.
• definition (fallback) — raw definition JSON. Only for chart types the typed model does not cover (composition, revenueLtv) or advanced params with no typed field.

Field names are snake_case; they are validated server-side by Pydantic.

The one thing that goes wrong

The compiler validates structure, not your taxonomy. An event or property name that doesn't exist won't error — it returns a well-formed chart with empty data, which reads like a real answer of "zero".

Resolve names first with the taxonomy search/read capabilities advertised by the connected catalog. Read event, user, or group properties as appropriate and use the exact name and scope returned. Do not invent a property-reader name from server prose or copy parameters from an older schema.

Two workflows

Create: discover names with the current taxonomy capability → build the typed chart → call query_amplitude_data → render_amplitude_chart with the returned chartEditId to show it.

Modify or fork a saved chart: search_amp_entities to find the chart id → get_amplitude_charts with include: 'typed' → edit the returned object → call query_amplitude_data with that chart and chartId set to the saved chart id. Passing chartId links the edit to its parent, and the parent's params fill any gaps the typed model omits. Reading a real chart back as typed params is also the fastest way to see how this project spells things.

get_amplitude_charts has five modes on include: link (default; validates ids, returns URLs, doesn't run the chart), typed, definition (raw), data (runs it, max 3 ids), and guide (schema for the raw fallback). All except guide need concrete ids — resolve them via search_amp_entities or get_from_url first.

Building blocks

These are shared across every chart kind.

Condition — one filter (+ Filter by, or a segment condition):

{ "property": "platform", "op": "is", "values": ["iOS"], "scope": "event" }

op: is, is not, contains, does not contain, greater than, less than, set, is not null. Use contains for prefix/substring matching. For presence checks prefer set — is not null is rejected by some chart kinds (data_table segments) even though error hints suggest it; set works everywhere. scope: event, user, group, session, or derivedV2 for computed properties — take the scope from the taxonomy lookup rather than guessing. Group-scoped properties also need group_type (see below).

Event — { event, where[], group_by[] }. A composite event puts several events in one slot: add object_type as INLINE_CUSTOM_EVENT (any member counts) or COMPARISON_EVENT (members compared), plus members[].

Segment — a population, combining property conditions and behaviors:

{
  "where": [{ "property": "country", "op": "is", "values": ["United States"], "scope": "user" }],
  "performed": [{ "event": "Purchase", "op": ">=", "count": 1,
                  "time_type": "rolling", "time_value": 30 }]
}

Omit segments entirely for all users. In performed, time_type defaults to forEachInterval; use rolling with time_value as lookback days.

Date range — required, and either relative or absolute, never both:

{ "relative": "Last 30 Days" }
{ "start": "2026-01-01", "end": "2026-01-31" }   // ISO 8601, or epoch seconds
{ "start": "now-90d" }                           // omit end for "up to now"

Optional timezone is an IANA name; omit for the project default.

The relative range's unit must match the interval: with weekly interval write "Last 12 Weeks", monthly "Last 6 Months" — "Last 3 Years" with a weekly interval is rejected. Re-denominate the window in the interval's unit or change the interval to match.

Interval — a word, not a number: hour, day (default), week, month, quarter. Sub-daily intervals only allow short windows (hour caps around 8 days).

count_unique_by — the counting entity, "User" by default, or a group like "org id".

Group-scoped properties

Group properties describe the account, not the user or the event — plan tier, ARR, industry, on the org id (or similarly named) group type. They are how almost every B2B question gets asked, and they are the most common thing to get stuck on.

Discover them explicitly with the connected taxonomy reader's group-property selector—they are not in the event or user catalogues.

Then use scope: "group" with the group_type the lookup returned, in a where filter or a group_by:

{ "property": "plan", "op": "is", "values": ["Enterprise"],
  "scope": "group", "group_type": "org id" }

Three rules that avoid nearly all of the failures seen in production:

1. group_type is the group's display name ("org id", "company") — exactly as the group-type and taxonomy readers spell it, lowercase and spaced. Not the property name, not "Group", not a grp:-prefixed key.
2. Never prefix the property name. Write "plan", not "grp:plan" or "gp:org id:plan" — those are storage-layer spellings that the chart API rejects.
3. Counting by account is a different setting. To count organizations rather than users, set count_unique_by: "org id". That is independent of whether you filter or break down by a group property.

When a group property is rejected

Invalid group property <name> for group type <type> means the backend's property registry has no key for that name — the property is advertised by the taxonomy API but not queryable. This is a data-plane gap, not a mistake in your query, and retrying spelling variants will not fix it. Retry loops burning ten-plus calls on this are the single largest source of wasted turns on this tool.

Do this instead, in order:

1. Re-read the exact name with the active taxonomy reader's group-property selector. If it differs from what you sent, correct it and retry once.
2. If it matches, stop retrying. Check for a user-scoped equivalent (accounts are often mirrored onto users, e.g. a user property plan) and use scope: "user".
3. If there is no equivalent, say plainly that the group property is not queryable in this project and give the user the answer you can produce — usually the same chart with count_unique_by: "org id" and no group breakdown.

Chart kinds

kind discriminates the union. All kinds take group_by, segments, date_range, interval, count_unique_by, and name (always set a descriptive name — it becomes the title).

segmentation

Needs events (at least one). measured_as.as defaults to unique_users; other values are event_totals, active_pct, avg_per_user, frequency, formula, histogram, and the property aggregations property_sum, property_avg, property_min, property_max, property_median, property_count, property_count_avg.

Property aggregations and histogram require the aggregated property to also appear in that event's group_by — omitting it is a common cause of an empty or malformed result.

For formula, put the expression in measured_as.formula using UPPERCASE functions (UNIQUES, TOTALS, PROPSUM, PERCENTILE, …) and refer to events as A, B, C in the order they are listed.

Also available: rolling_window (days), cumulative, period_over_period (period, parent, grandparent, quarter), and vis (line, bar, area, stackedbar, pie, kpi).

funnel

Needs steps (at least two). conversion_window is {value, unit} where unit is second, minute, hour, day, or week; a bare number or "7 days" also works, and omitting it gives the product default of 30 days. Set it explicitly whenever the request implies a window.

mode is ordered (default), unordered, or sequential. measured_as.as is conversion (default), conversion_over_time, time_to_convert, time_to_convert_over_time, or step_count.

constant_properties forces values to match across steps — this is how you build a same-session funnel (hold session_id). excluded_events takes {event, step_index}, where -1 excludes globally across all steps.

retention

Needs start_event (use _new for new users) and return_events (OR-combined; use _active for any activity). retention_method is rolling (default; on or after day N), nday (exactly day N), bracket (custom, via retention_brackets like [[0,7],[7,14]]), or nday_or_before. measured_as is retention or usage_interval.

sessions

measured_as is one of totalSessions (default), average, length, peruser, totalTime, averageTimePerUser, averageEventsPerSession, totalEvents, eventCountDistribution, or formula. Use groups for multiple series; omit it for a single unfiltered series.

data_table

Needs columns (at least one), each with a metric_type — UNIQUES, TOTALS, FORMULA, SESSIONS, CONVERSION, or the property aggregations PROPSUM, PROPAVG, PROPMAX, PROPMIN. rows are the breakdown dimensions, each kind either property or time.

Worked example

Weekly unique users of a signup event, on iOS, among users in the US, broken out by plan:

{
  "kind": "segmentation",
  "name": "iOS signups by plan",
  "events": [{ "event": "Sign Up", "where": [
    { "property": "platform", "op": "is", "values": ["iOS"], "scope": "event" }
  ]}],
  "measured_as": { "as": "unique_users" },
  "group_by": [{ "property": "plan", "scope": "user" }],
  "segments": [{ "where": [
    { "property": "country", "op": "is", "values": ["United States"], "scope": "user" }
  ]}],
  "date_range": { "relative": "Last 12 Weeks" },
  "interval": "week"
}

Compile errors

CompileChart failures come back as a 400 naming the offending field with a fix-oriented hint — read the hint, fix that one field, and retry. Do not rebuild the whole chart. The ones seen most in production:

1. Range/interval unit mismatch — Invalid range format: Last 3 Years … Match the relative range's unit to the interval. The message now carries the equivalent window for your interval; paste it in. (Last 30 Days at a weekly interval → Last 4 Weeks.)
2. Invalid group property — a registry gap, not a spelling mistake. See "When a group property is rejected" above; do not retry variants.
3. Unknown operator — the valid operator list differs per chart kind; the fix hint in the message is usually right, but for presence checks use set (works in every kind) rather than is not null.

Shape mismatches no longer need a retry: a bare string is accepted wherever a single-key wrapper object is expected and vice versa, so measured_as: "event_totals", events: ["Page Viewed"], and group_by: ["country"] all compile. Two consecutive failures on the same field means the field is not the problem — re-check the taxonomy, or tell the user what is blocking you rather than trying a third spelling.

When results come back empty

1. Re-read every event and property name with the current taxonomy search/reader and check the scope matches what it returned, including the group-property surface for account-level properties.
2. Widen date_range — the window may predate instrumentation.
3. Drop segments, then where, to find which filter empties the result.
4. For property aggregations, confirm the property is in the event's group_by.
5. Confirm the event is still arriving with check_for_recent_event_ingestion.