dotplot-mcp
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@dotplot-mcp Analyze events.csv and find my aha moment"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Dot Plot MCP
See individual users, not aggregate charts.
English | 한국어

Install
Paste this to your agent and it sets everything up:
Set up Dotplot (MCP + skills) — instructions are here https://dotplot-reports.vercel.app/setup.mdclaude mcp add --scope user dotplot -- uvx dotplot-mcpRelated MCP server: pendo-mcp
Use
Say this in any project:
"Analyze my product"
You get the report above. That's it.
Or use the slash commands:
/dotplot-analyze-product full analysis and report
/dotplot-add-tracking find and write the logging you're missing
/dotplot-whats-changed compare with the previous reportClaude finds your data, picks the action that means "this user got value", and
writes the report. No events table needed — your orders table already is one:
SELECT user_id, created_at::date AS date, 'purchase' AS event FROM orders
UNION ALL
SELECT user_id, added_at::date, 'add_to_wishlist' FROM wishlist_itemsNothing tracked yet? Say "add the tracking I'm missing" and Claude reads your code, writes the logging that's absent, and tells you when to come back.
git clone https://github.com/brownglasses/dotplot-mcp && cd dotplot-mcp
uv run sample_data.py # 40 fake users with a pattern planted in them
uv run harness.py # watch the whole pipeline runWhy this exists
DAU/MAU charts go "up and to the right" as long as new users arrive — even when nobody stays. Until you have hundreds of users, the most informative dashboard is YC's dot plot (David Lieb): one row per user, one cell per day.
Four rules keep it honest:
Code computes, AI only interprets — same data, same numbers, every time
Small samples withhold judgment — under 5 users in a group, it says nothing
Correlation isn't cause — every finding ships with "test this before you believe it"
Vanity metrics are refused — pick
open_appas your value event and the code says no
Reports come out in your language (English, 한국어, 日本語 built in; anything else translated on the fly with the numbers verified intact).
More
MIT
Available Tools
16 toolsanalyzeA
START HERE. Event data in, finished report out, one call.
Use this whenever the user asks anything general — "analyze my product", "how are my users doing", "find my aha moment". The other tools are parts; this is the whole thing. Only reach for them when the user asks for one specific number ("just show me retention").
It picks a value event, draws the dot plot, finds the aha moment, builds the funnel and retention curve, and writes the HTML report — then tells you what it found so you can say it out loud.
csv_path: a CSV with user_id, date, event (platform optional). No CSV yet? Call with no arguments and follow the instructions you get back — for a project with a database you will explore its schema and turn ordinary business tables (orders, sessions, posts) into events with load_from_db. Most early products have no events table; that is expected.
value_event: the action that means "this user got real value".
Leave it out and the code picks the candidate most users repeat. Pass it
yourself when you have read the codebase and know better — you can tell
purchase from view_item and the code cannot. The result always names
what was chosen and what else was available, so you can call again with a
different one if the choice looks wrong.
lang: the user's language. en/ko/ja are built in; for any other language call get_report_strings, translate, and use generate_report directly.
Inputs: lang, csv_path, output_path, value_event.
audit_trackingA
Check what the code logs, and — if there is data — whether it arrives.
Use this when analyze says there is nothing to work with. A project that tracks nothing has no data to analyse, but it has code to read, so leave csv_path out and this still answers.
How to use it:
Search the codebase for logging calls yourself. Common shapes: logEvent(...), track(...), analytics.capture(...), posthog.capture(...), gtag('event', ...), mixpanel.track(...)
Pass the event names you found as code_events.
Read the result:
in_code_never_fired logging is broken, or nobody uses that feature
in_data_not_in_code dead code, or your search missed it (look again)
Then find what has no logging at all — button handlers and core actions that should be recorded and aren't. That gap won't appear in either list, and it is usually the important one.
Prescribe, don't just report. For each hole write the one line of logging that belongs in that file, in that function, matching the surrounding style, show it, and ask whether to add it. Follow the project's existing naming (follow_artist if it is snake_case, followArtist if camelCase).
Once logging is in, tell them when to come back. come_back_in_days is in the result and is not a guess — it is how long the code must watch a user before it can honestly call them churned or a regular. Saying "run this again tomorrow" would produce a report with nothing in it.
Inputs: code_events (required), csv_path.
classify_usersC
Sort users by behaviour: churned (used once, never returned), weekend_only, regular (almost daily), casual.
lang sets the human-readable labels (en/ko/ja); the keys stay in English.
Inputs: csv_path (required), value_event (required), lang.
compare_benchmarkB
Compare your numbers with percentiles from teams at the same industry and stage.
Below ten teams the percentiles mean nothing — say so plainly instead of reporting them.
Inputs: stage (required), csv_path (required), industry (required), value_event (required).
describe_eventsA
Shape of an event CSV: date range, user count, events by type.
Mostly useful when you want to look before choosing a value event yourself. For a normal "analyze my product" request call analyze instead — it does this step and everything after it.
Inputs: csv_path (required).
dot_plotC
Text dot plot: one row per user, one cell per day.
◎ first active day, ● value event, · nothing. mark_events puts a letter on other actions, e.g. {"create_playlist": "P"}.
Inputs: csv_path (required), value_event (required), lang, weeks, +1 more. Returns: result
find_aha_momentsB
Scan every action for the one that turns users into regulars.
Ranked by behaviour_change — how much a user's own activity rose after they first did it — because comparing groups (lift) rewards actions that frequent users happen to do. Tell the user this is correlation, not cause.
Inputs: csv_path (required), value_event (required).
find_similar_casesB
Find companies that hit the same problem and what they changed.
A small library curated from public material — YC talks, First Round, founder interviews. When you pass a case on:
explain matched_on, so the user knows why it came up
cite the source, and say the figures are second-hand
offer to apply the fix to their code (e.g. move the aha action into onboarding) industry: b2c|b2b|commerce|content|social|tool|game|other — a match in the same industry ranks higher.
Inputs: csv_path (required), value_event (required), industry.
generate_reportA
Write the HTML dot plot report — readable in three seconds, made to share with a team or an investor.
analyze calls this for you. Use it directly when you need to control the marks, the window, or a language that isn't built in.
Marks on other actions:
default: the top aha events are picked automatically (behaviour change of 30 points or more, at most two, so the report stays quiet no matter how many event types exist)
your own: mark_events={"create_playlist": "P"}, described by mark_labels
none at all: mark_events={}
Language: match the conversation. lang="en"|"ko"|"ja" are built in; for any other language translate get_report_strings() and pass it as custom_strings with lang="custom". Never translate the {brace} placeholders.
Inputs: csv_path (required), value_event (required), lang, weeks, +5 more. Returns: result
get_report_stringsA
Every sentence the report can contain, in English, for translation.
For a language other than en/ko/ja:
call this,
translate the values — leave {n}, {rate_did} and every other brace placeholder exactly as they are, that is where the numbers go,
pass the result to generate_report as custom_strings with lang="custom". The code checks the placeholders survived, so the statistics stay exact.
Inputs: no named parameters.
history_compareA
Compare today's numbers with the last analysis of the same data.
Snapshots are saved to ./.dotplot/history.json every time a report is made. Only snapshots of the same dataset are compared, so the first run of a new project has nothing to compare against — say so rather than implying zero change.
Inputs: csv_path (required), value_event (required).
load_from_dbA
Pull events out of the project's database into a CSV.
How to use it:
The connection string comes from DOTPLOT_DB_URL by default. Tell the user to export it — never ask them to paste a password into the chat: export DOTPLOT_DB_URL="postgresql://readonly:...@host:5432/db" (Supabase: Dashboard > Settings > Database > Connection string)
Read the schema first (information_schema), find the tables that record what users did, and shape them into user_id, date, event. Most products have no events table — ordinary business tables are the event log: SELECT user_id::text, created_at::date AS date, 'purchase' AS event FROM orders Combine several actions with UNION ALL.
Only SELECT runs; anything else is refused. Recommend a read-only role.
Supports postgresql:// (Supabase, RDS, Neon) and sqlite:///path for testing.
Inputs: query (required), db_url, output_csv.
onboarding_funnelB
Onboarding funnel: signup -> first value -> came back -> still active.
Answers where new users leak. A large median_days_to_value means friction between signing up and getting anything out of the product.
Inputs: csv_path (required), value_event (required).
publish_reportA
Host the report and return a shareable link.
ASK THE USER FIRST. This puts their product's numbers on the public web. Only run it when they have said they want a link.
The report goes to a random path in a Vercel project and is deployed with
vercel deploy --prod. The random path plus a noindex header means only
someone with the link can reach it and search engines won't list it — but it
is still the open internet. Requires the vercel CLI to be logged in. To take
one down, delete the file under r/ and deploy again.
Inputs: html_path (required).
retention_curveB
Weekly retention: share of users with a value event in week N.
Only users old enough to have reached week N count toward it, and weeks with fewer than five of them are dropped rather than reported as noise. Where the curve flattens is the retention that holds.
Inputs: csv_path (required), value_event (required), max_weeks.
submit_benchmarkA
Submit five aggregate numbers to the anonymous benchmark. Opt-in.
Set consent=True only after the user has explicitly agreed. When you ask, show them exactly what leaves the machine — user count, churn rate, weekend rate, regular rate, aha lift, and nothing else. No user IDs, no event log, no product name.
industry: b2c|b2b|commerce|content|social|tool|game|other stage: pre_launch|under_100_users|under_1k_users|over_1k_users
Inputs: stage (required), csv_path (required), industry (required), value_event (required), +1 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 16 tools
Each tool has a distinct, clearly-stated purpose, and the descriptions explicitly disambiguate the large overlap between the meta-tool analyze and its component tools (dot_plot, find_aha_moments, retention_curve, generate_report by naming analyze as 'START HERE' and the others as 'parts'. Boundaries are clear in text but the pipeline-vs-composite overlap still requires careful reading to avoid picking a component when the whole workflow is wanted.
Names are uniformly snake_case, which is the key consistency factor. There is mild variation between verb_noun forms (get_report_strings, generate_report, classify_users, find_aha_moments) and noun phrases (dot_plot, retention_curve, onboarding_funnel), but the convention is predictable enough to navigate.
16 tools is slightly on the heavy side but justified for a full product-analytics suite spanning ingest, analysis, benchmarking, publishing, and history. Nearly every tool earns its place, though some (compare_benchmark, submit_benchmark, find_similar_cases) are peripheral.
The surface covers the lifecycle well: data ingestion (load_from_db, describe_events), the full analysis workflow (analyze plus components), external context (compare_benchmark, find_similar_cases, history_compare), and output (generate_report, publish_report). No obvious dead ends, though there is no raw-data export or report deletion tool.
Maintenance
Related MCP Connectors
Clamp Analytics MCP server: traffic, revenue, funnels, cohorts, errors, and search, for AI agents.
MCP server for querying and analyzing data from ad platforms, analytics tools, and spreadsheets
Product analytics, user and company insights, surveys and NPS, reports, dashboards, and more.
Analytics your AI agent can actually use. Track, experiment, and optimize via MCP.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceAn MCP server for data analysis and visualization supporting CSV and Excel files. It enables users to generate statistical summaries and create multi-dimensional charts like heatmaps and bar plots through natural language.-
- FlicenseNot gradedqualityDmaintenanceA comprehensive MCP server that provides AI assistants with powerful access to Pendo analytics through 15 specialized tools, transforming natural language questions into deep product insights.-
- AlicenseAqualityDmaintenanceFirst-party web analytics MCP server for AI agents, providing 42 tools to query traffic, events, funnels, conversions, sources, and performance data.4045 npmMIT
- AlicenseNot gradedqualityBmaintenanceMCP server that provides product and web analytics capabilities, allowing AI coding assistants to query funnels, retention, trends, feature flags, A/B tests, and more in natural language.5MIT