Terminal
Terminal/Skill
Guided modules that teach you how to use the Mixpanel MCP server, from first prompts to dashboards and governance.
By Mixpanel
Works with
npx skills add mixpanel/ai-plugins --skill learn-mcpEngine required — this skill is about the Mixpanel MCP server; if it's not set up, offer to run
/mixpanel:installfirst. If the resolved engine isheadless, point them to https://docs.mixpanel.com/docs/mixpanel-headless and exit.
Onboards users to the Mixpanel MCP server through interactive modules. One concept, one prompt, one result at a time.
Starter path (get value fast):
Advanced path (deeper investigation and governance):
Reference (open anytime):
If you're new to MCP, start with Module 0. Reply with a module name or number to begin.
Concept: MCP isn't a faster report builder. If you're only using it to skip clicks, you'll miss the point. The real unlock is exploration and synthesis: asking questions in plain language and joining product behavior with context that doesn't live in Mixpanel.
The one rule to internalize: for analysis prompts, include four things: Behavior (which events), Population (who), Timeframe (when), and Shape (rate, trend, breakdown). When any part is missing, the AI fills in defaults that may not match intent.
Want the deeper prompting principles, or ready to jump into Module 1 and orient your data?
═══════════ DEFERRED ═══════════
Surface the two or three that fit what they're doing:
Checkout Started to Purchase Completed for first-time buyers, last 60 days" beats "show me checkout conversion."Four phases: Discover (projects, events, properties) → Query (insights, funnels, retention, flows) → Create (dashboards, saved metrics, Lexicon edits) → Iterate (follow-ups in the same conversation). Query results are temporary — persist them by building a dashboard.
Concept: Map the project's schema before running any analysis. Schemas drift (event names get cryptic, properties go undocumented), and skipping this step is the most common reason MCP answers come back confidently wrong.
Try this:
"What Mixpanel projects do I have access to? Pick the most active one and list the top 10 events by recent volume, with descriptions if available. Flag any that don't have a description."
═══════════ DEFERRED ═══════════
A good result: project names and IDs, an event list with volume signals, descriptions where they exist and flags where they don't, and a read on what the product looks like.
Show the contrast — make the AI reason about the schema first:
"List the events that relate to checkout. For each, give the description and top three properties. Then suggest which combination best represents
purchase intentbased on how this project is actually instrumented."
Concept: This is where MCP replaces dashboard navigation. A well-formed prompt returns a chart and a written takeaway in one turn. Apply the four-part rule from Module 0 to every analysis prompt.
Try this:
"Pick a key funnel in this project. Run the conversion analysis for the last 30 days, break it down by one meaningful property. Tell me where the biggest drop-off is and which segment converts best."
═══════════ DEFERRED ═══════════
Show the contrast. The good version names all four parts:
"What's the conversion from
Sign UptoFirst Purchasefor users acquired through paid channels over the last 60 days? Show the trend week-over-week and flag any week the rate dropped more than 10%." Behavior (two events), population (paid-channel users), timeframe (60 days), shape (weekly trend with anomaly flags).
Concept: Some questions are unanswerable in aggregate. "Why did this account churn?" is a user-level question. This module drops from population data to individual users via Session Replay or Flows, depending on your instrumentation.
Try this:
"Identify 5 users who started but didn't complete a key funnel in the last 7 days. If Session Replay is enabled, pull their replays. If not, show the most common paths after the drop-off point and tell me the top 3 things they did instead of converting."
═══════════ DEFERRED ═══════════
"For users who dropped between
[Step A]and[Step B]in the last 14 days, show the most common paths after[Step A]. What are the top 3 things they did instead of converting?"
Show the two-step contrast — identify the users, then investigate:
"Step 1: find 10 users who finished onboarding in the last 30 days but haven't been active in 14. Step 2: pull replays for 3 of them, focused on their last active sessions, and summarize what they did before going quiet."
Concept: A query result lives only in the chat. To make an analysis persist, you turn it into a dashboard. Saved metrics also write back and can be reused across reports — a definition like "activated account" lives in one place instead of being rebuilt each time.
Try this:
"Take the analysis we've done and turn it into a Mixpanel dashboard. Pick a clear name. Add a text card at the top explaining what it tracks and how to read it."
═══════════ DEFERRED ═══════════
"Create a saved metric for [the core behavior we analyzed]. Show me the definition before you save it."
Concept: MCP can audit your Lexicon and fix what it finds, all from chat. Note: reading and auditing works for any role, but writing changes (descriptions, tags, merges, sensitive flags) requires Project Owner or Admin. The rule that keeps it safe: audit first, review the plan, then apply selectively — never let it write before you've seen what it will do.
Try this:
"Run a read-only Lexicon health check: list events missing a description, properties that look like PII but aren't flagged sensitive, and any duplicate event names. Don't change anything, just group findings by issue type."
═══════════ DEFERRED ═══════════
Group findings by type. Show the full plan for each group with a single confirmation per group:
"Draft a description for each undocumented event. Show me the full list, then apply only the ones I approve." "Mark these properties as sensitive: [list]." "Merge the duplicate cluster into [canonical name]." (reversible in the Mixpanel UI)
Lead with the count, not the raw list:
"How many distinct events have fired in the last 90 days, versus the total number defined?"
Concept: The highest-leverage thing MCP does is combine Mixpanel with your other connected tools for synthesis that used to take a week of manual coordination. The discipline that makes it work: decompose into one tool per turn, quantitative first, qualitative last.
Try this (follow these steps in order):
- Run a Mixpanel analysis that surfaces a finding worth investigating (a drop-off, an anomaly, a segment difference).
- Pick one other connected tool (error monitoring, CRM, tracker, feedback).
- Run a follow-up against that second tool to deepen the Mixpanel finding — e.g. correlate the drop-off with error spikes, or look up churned accounts in CRM.
═══════════ DEFERRED ═══════════
A feature shipped 3 weeks ago; leadership wants to know if it landed. One prompt per turn:
Concept: This module is reference, not a lesson. Pull from it when a question comes up. Ask which limit or principle they're checking on — don't recite the whole list.
═══════════ DEFERRED: surface only the one that applies ═══════════