This skill should be used when the user asks about Mixpanel product analytics, event data, funnel analysis, retention curves, cohort analysis, segmentation queries, user behavior, conversion rates, churn, DAU/MAU, ARPU, revenue metrics, feature adoption, A/B test results, user paths, flow analysis, session replay or session recordings (what a specific user did on screen, click-by-click — rage clicks, dead clicks, error sessions, action timelines), or any request to query, explore, visualize, or analyze Mixpanel data using Python. Also use when the user asks to read, write, or manage Mixpanel "business context" — the markdown documentation that grounds AI assistants in an organization's structure and goals.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
This skill should be used when the user asks about Mixpanel product analytics, event data, funnel analysis, retention curves, cohort analysis, segmentation queries, user behavior, conversion rates, churn, DAU/MAU, ARPU, revenue metrics, feature adoption, A/B test results, user paths, flow analysis, session replay or session recordings (what a specific user did on screen, click-by-click — rage clicks, dead clicks, error sessions, action timelines), or any request to query, explore, visualize, or analyze Mixpanel data using Python. Also use when the user asks to read, write, or manage Mixpanel "business context" — the markdown documentation that grounds AI assistants in an organization's structure and goals.
allowed-tools
Bash Read Write WebFetch
mixpanel_headless API Reference
Analyze Mixpanel data by writing and executing Python code using the mixpanel_headless library and pandas.
import mixpanel_headless as mp
ws = mp.Workspace()
result = ws.query("Login", last=30)
print(result.df.head())
Query Engines
Question
Method
Returns
How much? How many? Trends?
ws.query()
QueryResult
Do users convert through a sequence?
ws.query_funnel()
FunnelQueryResult
Do users come back?
ws.query_retention()
RetentionQueryResult
What paths do users take?
ws.query_flow()
FlowQueryResult
Who are they? What do they look like?
ws.query_user()
UserQueryResult
All result types have a .df property returning a pandas DataFrame and a .params dict containing the bookmark JSON.
FlowQueryResult also has .graph (NetworkX DiGraph) and .anytree (list of tree roots).
Guessing event names causes silent empty results. Guessing API parameters causes TypeErrors and invalid queries. Discover both the data schema AND the API surface before writing any query.
Step 1: Discover the data schema
import mixpanel_headless as mp
from mixpanel_headless import Filter, GroupBy, Metric
ws = mp.Workspace()
# 0. (Best first move) Map the WHOLE schema in one call — which properties are on# which events, plus per-property coverage. Ground yourself here so you never# filter or group by a property an event doesn't actually have.
schema = ws.schema_graph(include_density=True)
print(schema.properties_for_event("Login")) # exact property names on this eventprint(schema.relationships_df.head()) # event | property | density_local# 1. Find real event names
events = ws.events()
top = ws.top_events(limit=10)
print("Events:", events[:20])
print("Top:", [(e.event, e.count) for e in top])
# 2. Find real property names for the event you'll query
props = ws.properties("Login") # use an actual event name from step 1print("Properties:", props)
# 3. (Optional) Check property values to validate filter inputs
vals = ws.property_values("platform", event="Login")
print("Platforms:", vals)
Step 2: Discover the API surface with help.py
help.py is the source of truth for method signatures, parameter names, type constructors, and enum values. The method signatures later in this document are summaries — always verify with help.py before using a method or type you haven't looked up.
Never guess parameter names. If you're unsure whether a parameter is called property or math_property, or what arguments GroupBy() accepts, run help.py first. Wrong parameter names cause TypeErrors that waste tool calls.
# Look up a query method BEFORE writing the query
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py Workspace.query
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py Workspace.query_funnel
# Look up types BEFORE constructing them
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py Filter # → classmethods: .equals(), .less_than(), etc.
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py Metric # → property=, NOT math_property=
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py GroupBy # → property, property_type only
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py MathType # → enum values# Look up result types to know what columns .df returns
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py QueryResult
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py FlowQueryResult
# Search when you're not sure of the exact name
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py search cohort # → CohortBreakdown, CohortMetric, CohortDefinition, ...
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py search retention # → query_retention, RetentionEvent, RetentionMathType, ...# List everything
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py types # all public types
python3 ${CLAUDE_SKILL_DIR}/scripts/help.py exceptions # all exceptions
For tutorials and guides: WebFetch(url="https://mixpanel.github.io/mixpanel-headless/llms.txt")
Discovery method signatures
defevents(self) -> list[str]: ...
# List all event names (cached).defproperties(self, event: str) -> list[str]: ...
# List all property names for an event (cached).defproperty_values(self, property_name: str, *, event: str | None = None, limit: int = 100) -> list[str]: ...
# Get sample values for a property.deftop_events(self, *, type: Literal['general', 'average', 'unique'] = 'general', limit: int | None = None) -> list[TopEvent]: ...
# Get today's most active events. TopEvent has .event (str), .count (int), .percent_change (float).deffunnels(self) -> list[FunnelInfo]: ...
# List saved funnels.defcohorts(self) -> list[SavedCohort]: ...
# List saved cohorts.deflist_bookmarks(self, bookmark_type: BookmarkType | None = None) -> list[BookmarkInfo]: ...
# List all saved reports (bookmarks).deflexicon_schemas(self, *, entity_type: EntityType | None = None) -> list[LexiconSchema]: ...
# List Lexicon schemas (event/property definitions).defschema_graph(self, *, include_density: bool = False, include_user_properties: bool = True, force_refresh: bool = False) -> SchemaGraphResult: ...
# Whole-project Lexicon + event<->property graph in one call. SchemaGraphResult has# .relationships_df (event|property|density_local), .properties_for_event(e),# .events_for_property(p), .orphan_properties(), .to_graph() (networkx DiGraph).defclear_discovery_cache(self) -> None: ...
# Clear cached discovery results.# User Guide: WebFetch(url="https://mixpanel.github.io/mixpanel-headless/guide/discovery/index.md")
Exploratory Analysis Workflow
When exploring an unfamiliar dataset or asked to "find insights," follow this systematic approach. Do NOT skip to querying — explore first.
Step 1: Orient — Map the Event Schema
Start with schema_graph(). One call returns the whole event↔property map plus
per-property coverage (density_local), so you learn which properties actually
travel with which events — and how well-populated each is — before querying. This
is the single most useful grounding step: it stops you filtering or grouping by a
property an event doesn't carry.
import mixpanel_headless as mp
ws = mp.Workspace()
schema = ws.schema_graph(include_density=True)
print("events:", schema.meta["event_count"],
"| event properties:", schema.meta["event_property_count"])
# Headline view: one row per (event, property), with coverage.print(schema.relationships_df.head(20)) # event | property | density_local# Exact properties on each top event — use these names verbatim in queries.for event_name inlist(schema.event_to_properties)[:5]:
print(f"\n{event_name}: {schema.properties_for_event(event_name)}")
# Properties attached to no events — usually noise; skip them.print("\nOrphans:", schema.orphan_properties()[:20])
top = ws.top_events(limit=15)
print("\nTop by volume:", [(e.event, e.count) for e in top])
Then sample values for the properties you'll actually group or filter by:
for p in schema.properties_for_event("Purchase")[:15]:
print(p, ws.property_values(p, event="Purchase", limit=10))
Step 2: Classify Properties
Infer property types from sampled values to decide how to use each:
Boolean: values are ['true', 'false'] — segment with group_by, often pre-computed behavioral flags
Low-cardinality categorical (<10 values): platform, tier, category — use for group_by
Numeric: values parse as int/float: price, total, count — use for math='average' or math='sum' with math_property=
High-cardinality (>100 values): IDs, names — skip for group_by, may need custom property cleanup
Temporal: ISO dates or epoch values — use for time-based analysis
Property naming patterns that signal analytical value:
*_total, *_count, *_value, *_amount → numeric, aggregate with avg/sum/median
*_name, *_type, *_category, *_tier → categorical, use for breakdowns
*_id, *_uuid → identifiers, skip for breakdowns
Step 3: Scan for Significant Segments
For each boolean and low-cardinality categorical property on key events, run a quick breakdown against a numeric metric:
# Example: scan all interesting properties on a purchase event
numeric_prop = 'order_total'# or whatever the key metric is
interesting_props = [p for p in props ifnot p.endswith('_id')]
for prop in interesting_props:
vals = ws.property_values(prop, event=event, limit=10)
iflen(set(vals)) <= 10: # low cardinality — worth a breakdown
result = ws.query(event, math='average', math_property=numeric_prop,
group_by=prop, last=90, mode='total')
print(f"\n{numeric_prop} by {prop}:")
print(result.df.to_string(index=False))
# Flag segments where metric differs >15% from overall
Step 4: Deep Dive on Significant Findings
When a breakdown reveals a notable difference (>15% between segments):
Quantify: calculate the exact ratio between segments
Cross-reference: does this segment differ on other metrics too?
Investigate causally: run funnels or retention filtered by the segment
Control for confounds: add a second group_by dimension to check if the effect holds
Step 5: Analyze Messy String Properties
When string properties have complex/unreadable values (e.g., campaign names from tools like Braze):
Sample 15-20 values to identify the naming convention
Look for structural patterns: date codes, targeting prefixes, channel suffixes, audience tags
Design regex cleanup rules, one layer per structural element
Create a custom property with ws.create_custom_property(CreateCustomPropertyParams(...))
Verify by querying with the custom property as group_by
Custom Property Formula Reference
Formulas use a SQL-like expression language. Variables (A, B, _A, etc.) map to properties via composedProperties.
REGEX_EXTRACT(haystack, pattern, capture_group) — returns match or capture group
REGEX_REPLACE(haystack, pattern, replacement) — replaces all matches
Regex engine quirks (Mixpanel-specific):
Case-insensitive by default — use (?-i) to switch to case-sensitive matching within a pattern
Backreferences work — $1, $2 capture groups and $0 whole-match all work in REGEX_REPLACE replacements
{n,m} quantifiers conflict with formula syntax — curly braces are parsed as formula constructs. Use repeated character classes instead (e.g., [0-9][0-9][0-9][0-9] instead of [0-9]{4})
\d, \w shorthand classes don't work — use [0-9], [A-Za-z0-9_] explicitly
Escape backslashes carefully — in formula strings, \\\\ may be needed for a literal \ depending on how the formula is constructed (Python string → JSON → regex engine)
CamelCase splitting — insert space between lowercase→uppercase boundaries:
REGEX_REPLACE(text, "(?-i)([a-z])([A-Z])", "$1 $2")
// ChickenSundaysApril → Chicken Sundays April
Practical multi-step cleanup example (campaign names from Braze):
Same parameters as the corresponding query methods, but return dict[str, Any] bookmark params without making an API call. Useful for creating saved reports (bookmarks).
Every query engine has parameters that look like simple settings but are actually analytical choices with outsized influence on results. Before running any query, apply these principles:
Find the master dial. Each engine has one parameter (or small set) that reshapes all downstream metrics. Changing it changes the story. Know which parameter it is and choose deliberately — don't accept defaults blindly.
Match parameters to the domain. There are no universal "correct" values. A social app needs daily retention; a B2B SaaS needs monthly. A food-ordering funnel needs a 1-hour window; an onboarding funnel needs 14 days. The product's natural usage cadence dictates the setting.
Distrust averages. Averages include outliers — one extreme value distorts the whole metric. Use medians (math='median', percentile=50) to see what's typical. If the mean and median diverge, the distribution is skewed and the mean is misleading.
Counting methodology is a modeling choice. "Unique users," "total events," and "sessions" aren't just modes — they answer fundamentally different questions. "How many people?" vs "How much activity?" vs "How many engagement moments?" Choose the counting method that matches the business question.
Know the silent defaults. Parameters are sometimes silently ignored (e.g., math_property with math='unique'), silently constraining (e.g., no funnel re-entry by default), or silently inflating (e.g., unbounded_mode='carry_forward' in retention). If results look surprising, check whether a default is shaping them.
Sweep, don't guess. When unsure which parameter value to use, try several and observe how metrics shift. Where the metric stabilizes or diverges reveals the true signal. The code examples below demonstrate this for each engine.
Comparing Segments Across Multiple Dimensions
When a single breakdown shows a difference, verify it holds across dimensions:
# Step 1: Find the interesting segment
result = ws.query(event, math='average', math_property='order_total',
group_by='deal_sweet_spot', last=90, mode='total')
# Found: deal_sweet_spot=true has 37% higher AOV# Step 2: Check if it holds across another dimension
result = ws.query(event, math='average', math_property='order_total',
group_by=['deal_sweet_spot', 'platform'], last=90, mode='total')
# Does the sweet spot hold for both iOS and Android?# Step 3: Check segment rates across a third dimension
result = ws.query(event, math='unique',
group_by=['loyalty_tier', 'deal_sweet_spot'], last=90, mode='total')
# Which tier is most likely to achieve the sweet spot?
Insights Analysis: MathType, Per-User, and the Unit of Analysis
MathType is the most critical Insights parameter. It defines what you're counting — total (event volume), unique (user reach), dau/wau/mau (time-bounded engagement), average/median/percentile (property distributions), sum (revenue totals). Choosing the wrong MathType answers the wrong question silently. Match MathType to the business question: engagement → dau or wau; revenue → sum or average with math_property; adoption → unique; intensity → total.
Per-user aggregation is a two-stage process that fundamentally changes the unit of analysis. per_user='average' with math='average' first computes each user's average, then averages across users. This is NOT the same as a global average — a power user with 1000 events and a casual user with 2 events contribute equally. This is often the right choice (prevents power users from dominating) but it changes the story dramatically:
# Sweep MathTypes to understand an event from multiple angles
event = 'Purchase'# use a real event name
prop = 'order_total'# use a real numeric propertyfor mt in ['total', 'unique', 'dau', 'average', 'median', 'sum']:
kwargs = {'math': mt, 'last': 30, 'mode': 'total'}
if mt in ('average', 'median', 'sum'):
kwargs['math_property'] = prop
result = ws.query(event, **kwargs)
print(f"{mt:>10}: {result.df['count'].iloc[0]:>12,.2f}")
# total = event volume, unique = user reach, dau = daily engagement,# average/median = typical transaction, sum = total revenue# Per-user aggregation: compare global average vs per-user average
global_avg = ws.query(event, math='average', math_property=prop, last=30, mode='total')
per_user_avg = ws.query(event, math='average', math_property=prop,
per_user='average', last=30, mode='total')
print(f"Global avg: {global_avg.df['count'].iloc[0]:.2f}")
print(f"Per-user avg: {per_user_avg.df['count'].iloc[0]:.2f}")
# If these differ significantly, power users are skewing the global average
Prefer medians over averages for property distributions — same principle as funnel time-to-convert. Use math='median' instead of math='average', or math='percentile' with percentile_value=50. Averages include outliers; medians reveal what's typical.
Silent traps:math_property is silently ignored with non-property MathType (e.g., math='unique' discards math_property). rolling reduces data point count without warning (a 30-day rolling window over 59 days produces ~29 points, not 59). unit is silently ignored in mode='total'.
Funnel Analysis: Windows, Modes, and Time
Funnel queries return time data in step metadata columns (avg_time, avg_time_from_start) alongside conversion rates.
Conversion window is the most critical funnel parameter. It defines the maximum time a user has to complete the funnel from their first step. It affects every other metric — conversion rate, time-to-convert, and segment comparisons all shift dramatically with window size.
Choosing a window: Match it to the user journey being measured. Short funnels (ordering food, adding to cart) need tight windows — hours, not days. Long funnels (onboarding, subscription purchase) need wider windows — days or weeks. When unsure, experiment:
# Try progressively tighter windows to find where signal emergesfor window, unit in [(14, 'day'), (7, 'day'), (1, 'day'), (12, 'hour'), (6, 'hour')]:
result = ws.query_funnel(steps, last=90,
conversion_window=window, conversion_window_unit=unit)
final = result.df[result.df['step'] == result.df['step'].max()].iloc[0]
print(f"{window}{unit}: conv={final['overall_conv_ratio']:.3f} "f"time={final['avg_time_from_start']/3600:.1f}h")
# Look for: conversion stabilizing, time differences appearing at tighter windows
Conversion counting modes change what "conversion" means:
conversion_rate_unique (default): unique users who completed. No re-entry — first attempt in the window or out.
conversion_rate_total: total completions. One user can count multiple times.
Combine with reentry_mode='basic' or reentry_mode='optimized' for multiple attempts. Optimized re-entry picks the best completion path.
Time-to-convert: prefer medians over averages. Average time includes outliers — one slow user inflates the average; one fast user pulls it down. Use median or percentiles for true speed trends:
# Median time via funnel math
result = ws.query_funnel(steps, last=90, math='median')
# Compare segments with filtered funnels + tight window
ios = ws.query_funnel(steps, where=Filter.equals('platform', 'iOS'),
last=90, conversion_window=6, conversion_window_unit='hour')
android = ws.query_funnel(steps, where=Filter.equals('platform', 'Android'),
last=90, conversion_window=6, conversion_window_unit='hour')
# Compare avg_time_from_start on matching steps
When comparing segments across funnels: always try at least 2-3 conversion windows. A difference invisible at 14 days may be stark at 6 hours. This is especially true for speed comparisons — tighter windows filter out noise and reveal which segment completes faster.
order changes what "conversion" means.'loose' (default) requires steps in sequence but allows other events between them. 'any' requires all steps in any order — a user who does C→B→A counts as converting. The difference is dramatic: loose funnels measure sequential workflows; any-order funnels measure feature adoption breadth. When unsure, run both and compare:
# order='loose' vs 'any' — same steps, fundamentally different questionsforordin ['loose', 'any']:
result = ws.query_funnel(steps, last=90, order=ord)
print(f"order={ord}: {result.overall_conversion_rate:.3f}")
# If 'any' >> 'loose', users are completing all steps but not in the expected order# This often reveals UX issues — users accomplish the goal but not via the designed path
holding_constant isolates cross-step consistency. Hold a property like 'platform' or 'device_id' constant and users who change values between steps (e.g., sign up on iOS, purchase on Android) are excluded. This reveals single-device vs cross-device conversion and is essential for understanding journeys that span platforms. Maximum 3 properties.
Exclusions disqualify tainted journeys.exclusions=["Logout"] removes users who logged out between funnel steps — unlike flow hidden_events, exclusions completely remove users from the funnel. Use for support escalation events, churn signals, or any action that taints the conversion path. Control which steps the exclusion applies to with Exclusion("Logout", from_step=0, to_step=2) (0-indexed).
Per-step filters narrow individual steps without affecting others.FunnelStep("Purchase", filters=[Filter.greater_than("amount", 50)]) restricts which Purchase events count, but doesn't filter Signup events. Global where filters ALL steps. This distinction is subtle but powerful: filter the population with where, filter the definition of a step with per-step filters.
Session windows are a distinct paradigm.conversion_window_unit='session' constrains the entire funnel to a single engagement session — no multi-session hops. This reveals true in-session conversion behavior, separate from users who spread a journey across days. The third counting mode, math='conversion_rate_session', counts sessions rather than users or events (requires conversion_window_unit='session').
Retention Analysis: The Cohort Bucketing Triple
retention_unit + alignment + bucket_sizes define your entire retention model — retention's equivalent of the funnel conversion window. retention_unit groups users into cohorts (day/week/month). alignment anchors cohorts (birth = each user's clock starts from their event; interval_start = snap to calendar boundaries). bucket_sizes sets measurement points. Changing any one reshapes all downstream metrics.
Match retention_unit to your product's natural usage cadence. Daily products (social, messaging) need retention_unit='day'. Weekly products (task management, fitness) need 'week'. Monthly products (subscriptions, B2B SaaS) need 'month'. When unsure, experiment:
# Sweep retention_unit to find natural product cadence
born, ret = 'Signup', 'Login'# use real event namesfor ru in ['day', 'week', 'month']:
result = ws.query_retention(born, ret, retention_unit=ru, last=90)
avg = result.average
if avg isnotNoneandlen(avg) > 1:
bucket_1_rate = avg.iloc[1]['rate'] if'rate'in avg.columns elseNoneprint(f"{ru:>6} retention: bucket 1 = {bucket_1_rate}")
# The unit where bucket-1 retention is highest reveals natural usage cadence# Custom buckets for milestone-based retention (days 1, 3, 7, 14, 30)
result = ws.query_retention(born, ret, retention_unit='week',
bucket_sizes=[1, 3, 7, 14, 30], unit='day', last=90)
print(result.df[result.df['cohort_date'] == '$overall'])
# Day 1 = activation, Day 7 = habit formation, Day 30 = long-term retention# Compare alignment modes — can shift results dramaticallyfor align in ['birth', 'interval_start']:
result = ws.query_retention(born, ret, retention_unit='week',
alignment=align, last=90)
print(f"\nalignment={align}:")
print(result.average.head() if result.average isnotNoneelse"No data")
Be wary of unbounded modes — they inflate retention.unbounded_mode='carry_forward' credits future returns to past buckets — a user who returns only on day 30 gets counted as retained in all buckets from 30 onward. carry_back inflates early buckets instead. Useful for "did they ever engage?" analysis but distorts standard retention curves.
retention_cumulative=True masks re-engagement gaps. Cumulative retention creates monotonically increasing curves where each bucket includes all prior buckets. This hides whether users who returned in week 1 ALSO returned in week 2. Standard (non-cumulative) retention reveals re-engagement patterns and true habit formation.
Counting methodology:math='retention_rate' (% who returned — the default), math='unique' (count who returned), math='total' (how many times they returned — events, not users). total reveals engagement intensity; a user logging in 5 times in bucket 1 counts as 5, not 1. Like funnels, the counting choice changes the question.
Flow Analysis: Windows, Cardinality, and Signal-to-Noise
Cardinality controls signal-to-noise — the most important flow-specific parameter. Low cardinality (2-3) reveals dominant paths — the main story. High cardinality (10+) reveals edge cases and niche journeys. Start low to find the narrative, then increase to find exceptions.
conversion_window matters for flows too — identical concept to funnels. Session-based windows (conversion_window_unit='session') reveal in-app behavior within a single engagement. Calendar windows reveal multi-day journeys. A tight window isolates intentional workflows; a wide window captures exploratory meandering:
# Sweep cardinality to find signal-to-noise sweet spot
event = 'Login'# use a real anchor eventfor card in [2, 3, 5, 10]:
result = ws.query_flow(event, forward=3, cardinality=card, last=30)
transitions = result.top_transitions(5)
print(f"\ncardinality={card}: {len(transitions)} top transitions")
for src, dst, count in transitions[:3]:
print(f" {src} → {dst}: {count}")
# Low cardinality = clear narrative; high cardinality = exhaustive but noisy# Compare count types (same principle as funnels)for ct in ['unique', 'total', 'session']:
result = ws.query_flow(event, forward=3, count_type=ct, last=30)
dropoff = result.drop_off_summary()
print(f"\n{ct}: step 0 dropoff = {dropoff}")
# unique = how many people; total = how much activity; session = how many sessions# Compare collapse_repeated to separate intent from noisefor collapse in [False, True]:
result = ws.query_flow(event, forward=3, collapse_repeated=collapse,
cardinality=5, last=30)
print(f"\ncollapse_repeated={collapse}:")
for src, dst, count in result.top_transitions(3):
print(f" {src} → {dst}: {count}")
collapse_repeated changes what "a path" means. With False (default), A→A→A→B is a distinct path from A→B — repetitive clicks look like distinct journeys. With True, consecutive duplicates merge, revealing intent over noise. Toggle this to see both the raw behavior and the simplified user intent.
hidden_events vs exclusions — hiding vs disqualifying.hidden_events removes events from display but they still affect path structure and counts. exclusions disqualifies users who performed those events entirely — a much stronger operation. Use hidden_events for decluttering (e.g., ubiquitous page views); use exclusions for removing tainted journeys (e.g., users who churned mid-flow).
Three modes reveal different stories.sankey shows aggregate flow structure and bottlenecks (where do most users go?). paths shows exact user journeys in sequence (what are the top 5 complete paths?). tree shows branching decision points (where do users diverge?). Use all three on the same data to build a complete picture.
User Profile Analysis: Modes, Aggregates, and Distribution Shape
mode is the most critical user query parameter.'profiles' returns individual user records (one row per user). 'aggregate' returns a single statistic. These are fundamentally different operations — profiles is a data extraction, aggregate is a calculation. Aggregate is also dramatically faster (single API call vs paginated fetching).
Sweep aggregate functions to understand distribution shape before building expensive profile queries. count tells you "how many." extremes reveals range (min/max). percentile at 50 gives median. numeric_summary gives mean, variance, and sum-of-squares:
# Sweep aggregate functions to understand a property's distribution
prop = 'lifetime_value'# use a real numeric profile propertyfor agg in ['count', 'extremes', 'percentile', 'numeric_summary']:
kwargs = {'mode': 'aggregate', 'aggregate': agg}
if agg != 'count':
kwargs['aggregate_property'] = prop
if agg == 'percentile':
kwargs['percentile'] = 50# median
result = ws.query_user(**kwargs)
print(f"{agg:>16}: {result.aggregate_data}")
# count = population size, extremes = range, percentile@50 = median,# numeric_summary = full distribution stats# If mean (from numeric_summary) >> median (from percentile), distribution is right-skewed# Point-in-time comparison with as_of
today_count = ws.query_user(mode='aggregate', aggregate='count',
where=Filter.equals('plan', 'premium'))
past_count = ws.query_user(mode='aggregate', aggregate='count',
where=Filter.equals('plan', 'premium'), as_of='2025-01-01')
print(f"Premium users: {past_count.value} (Jan 1) → {today_count.value} (today)")
Prefer medians over averages — same principle as funnels and Insights. aggregate='percentile', percentile=50 gives median; numeric_summary gives mean. If they diverge significantly, the distribution is skewed and the mean is misleading.
as_of enables temporal analysis — query profiles as they existed at a past date. Compare population states over time: "how many premium users existed on Jan 1 vs today?" Without as_of, you always see current state, making growth and churn invisible.
Inline CohortDefinition vs saved cohorts. Inline cohorts (cohort=CohortDefinition.all_of(...)) let you define complex behavioral segments on-the-fly without roundtripping to save/delete in Mixpanel. Much faster iteration for exploratory analysis. Use saved cohorts for production dashboards and monitoring.
Analytical Building Blocks: Custom Properties, Cohorts, and Frequency
Raw data is rarely analysis-ready. These three tools transform raw events and properties into analytically useful dimensions, populations, and segments. Recognize when to reach for each — they compose with every query engine.
Inline Custom Properties — transform data at query time. When property values are messy, need bucketing, or you need to derive new dimensions, create an InlineCustomProperty rather than querying raw values. Key patterns:
Bucketing continuous values for breakdowns (revenue → Low/Medium/High)
Cleaning messy strings with IFS/REGEX_EXTRACT (campaign names, UTM parameters)
Deriving new dimensions from arithmetic or date functions (profit margin, days since signup)
Fallback chains across multiple properties (display_name → username → "unknown")
from mixpanel_headless import InlineCustomProperty, PropertyInput, GroupBy, Filter, Metric
# Bucket revenue into tiers for breakdown
revenue_tier = InlineCustomProperty(
formula='IFS(A < 50, "Low", A < 200, "Medium", TRUE, "High")',
inputs={"A": PropertyInput("revenue", type="number")},
property_type="string",
)
result = ws.query("Purchase", group_by=GroupBy(property=revenue_tier), last=30, mode='total')
# Derive profit margin for aggregation
margin = InlineCustomProperty.numeric("(A - B) / A * 100", A="revenue", B="cost")
result = ws.query(Metric("Purchase", math="average", property=margin), last=30)
# Clean messy strings for segmentation
domain = InlineCustomProperty(
formula='REGEX_EXTRACT(A, "@(.+)$")',
inputs={"A": PropertyInput("email", type="string")},
property_type="string",
)
result = ws.query("Signup", group_by=GroupBy(property=domain), last=30, mode='total')
Use InlineCustomProperty for ad-hoc exploration. When a formula proves valuable, persist it with ws.create_custom_property() and reference it via CustomPropertyRef(id) across reports.
Inline Cohorts — define complex populations on-the-fly. Every analytical question starts with "among WHICH users?" Simple property filters (where=Filter.equals(...)) answer "users with attribute X." Inline cohorts answer harder questions: "users who did X at least N times in the last D days AND did NOT do Y AND have property Z." Compose criteria with AND/OR logic:
from mixpanel_headless import CohortDefinition, CohortCriteria, CohortBreakdown, CohortMetric
# "Power users": purchased 5+ times in 30 days, never contacted support
power_users = CohortDefinition.all_of(
CohortCriteria.did_event("Purchase", at_least=5, within_days=30),
CohortCriteria.did_not_do_event("Support Ticket", within_days=90),
)
# Use inline cohort as a breakdown — no need to save first
result = ws.query("Login", group_by=CohortBreakdown(power_users, "Power Users"), last=30)
# Use inline cohort as a filter in user queries
result = ws.query_user(cohort=power_users, mode='aggregate', aggregate='count')
# Track saved cohort size over time alongside event metrics
result = ws.query(
[Metric("Login", math="unique"), CohortMetric(saved_cohort_id, "Power Users")],
formula="(B / A) * 100", formula_label="% Power Users Active", last=90,
)
Frequency Breakdown/Filter — segment by behavioral intensity.FrequencyBreakdown answers "how do users who did X once differ from users who did X ten times?" FrequencyFilter restricts queries to users meeting a frequency threshold. These bridge "what users did" with "who users are":
from mixpanel_headless import FrequencyBreakdown, FrequencyFilter
# Break down login behavior by purchase frequency
result = ws.query("Login", math='unique',
group_by=FrequencyBreakdown("Purchase", bucket_size=3, bucket_min=0, bucket_max=15),
last=30, mode='total')
# Reveals: do frequent purchasers also log in more?# Filter to users who purchased 3+ times in 30 days, then analyze their flow
result = ws.query_flow("Login", forward=3,
where=FrequencyFilter("Purchase", value=3, date_range_value=30, date_range_unit="day"),
last=30)
# Reveals: what do repeat purchasers do after logging in?
When to reach for each:
Property values are messy or need derivation → Custom Property
Population requires behavioral criteria (did X, didn't do Y, frequency thresholds) → Inline Cohort
You need to segment by event frequency (how often, not just whether) → FrequencyBreakdown/Filter
You need to compare in-cohort vs out-of-cohort behavior → CohortBreakdown with include_negated=True
You need to track a segment's size as a time series → CohortMetric (saved cohorts only)
Legacy Queries & Counts
These use older APIs. Prefer the typed query methods above when possible.
All entity methods require a workspace ID. Use python3 ${CLAUDE_SKILL_DIR}/scripts/help.py Workspace.<method> for full signatures and parameter types.
User Guide: WebFetch(url="https://mixpanel.github.io/mixpanel-headless/guide/entity-management/index.md")
Data Deletion:list_deletion_requests, create_deletion_request, cancel_deletion_request, preview_deletion_filters
Other:get_tracking_metadata
Business Context
Read and write the markdown documentation that grounds AI assistants in your organization's structure and goals, exposed as a typed Python API.
Two scopes — level="organization" (shared across the whole org) and level="project" (per-project). 50,000-character cap enforced client-side before any HTTP call so oversize input fails fast. Org-level operations auto-resolve organization_id from the cached /me response; pass organization_id=N to override.
Run python3 ${CLAUDE_SKILL_DIR}/scripts/help.py search business_context to see all four methods, two types, and one exception.
from mixpanel_headless import BUSINESS_CONTEXT_MAX_CHARS # 50_000# Read
project_ctx = ws.get_business_context(level="project")
org_ctx = ws.get_business_context(level="organization") # auto-resolves org_id
explicit = ws.get_business_context(level="organization", organization_id=42)
# Read both at once (single round-trip via /business-context/chain)
chain = ws.get_business_context_chain()
print(chain.organization.content)
print(chain.project.content)
# Write (full-replace; pass "" to clear, or use clear_business_context())
ws.set_business_context("# About Acme\n…", level="project")
ws.set_business_context("# Org-wide standards", level="organization")
ws.clear_business_context(level="project")
# All return BusinessContext with: level, content, organization_id, project_id# Plus convenience .is_empty and .character_count properties (Python only)print(f"{project_ctx.character_count}/{BUSINESS_CONTEXT_MAX_CHARS} chars; "f"empty={project_ctx.is_empty}")
When to reach for this:
User asks "what's the business context for this project/org?" → get_business_context_chain()
User wants to version-control project context as a .md file → ws.set_business_context(Path("ctx.md").read_text(), level="project") in CI
User asks to "audit which projects have AI context configured" → iterate ws.projects() + ws.use(project=...) + ws.get_business_context(level="project") and check .is_empty
User asks to seed a new project from the org default → chain = ws.get_business_context_chain(); ws.set_business_context(chain.organization.content, level="project")
Permissions: project-scope reads need any project access; project-scope writes need edit_project_info on the project. Org-scope writes need edit_project_info at the org level (typically OAuth, not service account). The BusinessContextValidationError exception is raised client-side BEFORE any HTTP call when content exceeds 50,000 chars, so use it to detect oversize input without burning a round-trip.
User Guide: WebFetch(url="https://mixpanel.github.io/mixpanel-headless/guide/business-context/index.md")
Session Replay
Answers "what did this user actually do?" — the click-by-click story behind an analytics number. Fetches rrweb session recordings, runs a vendored analyzer, and projects sessions into DataFrames + an LLM-friendly action timeline.
When to reach for this: the user names a specific distinct_id and asks what they did, wants clicks / rage-clicks / error sessions across a cohort of sessions, or wants to correlate a tracked event with on-screen behavior.
import mixpanel_headless as mp
ws = mp.Workspace()
# Discover + fetch + Mixpanel-event join in one call → a ReplayBundle.# Each replay is byte-heavy, so limit defaults to 20 — raise it deliberately.
bundle = ws.replays_for_user("user-42", from_date="2025-01-01", to_date="2025-01-31")
bundle.sessions_df # one row per session: duration_s, n_clicks, n_errors, entry/exit_url
bundle.replays[0].summary_markdown # action timeline: "Clicked …", "Scrolled (×3)"
bundle.top_clicks(10) # most-clicked elements (focus interactions excluded)
bundle.rage_clicks() # rapid repeated clicks on one target
bundle.error_sessions() # a NEW bundle of only the replays with console errors
ReplayBundle projections: sessions_df, actions_df (includes a description column — the full phrase), events_df (raw rrweb), mixpanel_df (tracked events in the window), elements_df (per-element click counts, URL-normalized). Filters return new bundles: .where(distinct_id=, contains_url=, has_event=, min_duration_s=), .filter(predicate), .find_pattern([...]), .head(n), .sample(n, seed), .compare(other).
Single replay + raw stream for the rrweb JS player:
Signed CDN URLs are bearer credentials — SignedReplay masks them and the library never logs them. A SESSION_RECORDING_SENSITIVE_DATA 403 raises SessionReplayAccessError.
Look up the surface: help.py Workspace.replays_for_user, help.py ReplayBundle, help.py Replay.
User Guide: WebFetch(url="https://mixpanel.github.io/mixpanel-headless/guide/session-replay/index.md")
Key Types
Run python3 ${CLAUDE_SKILL_DIR}/scripts/help.py types for the full list of all types. Use help.py <TypeName> for fields, constructors, and enum values.
Full reference: WebFetch(url="https://mixpanel.github.io/mixpanel-headless/api/types/index.md")
