How it works

Your question gets read before anything gets built.

You write a question in your own words. Before anything runs, we read what kind of question it is, what field it comes from and what shape of data it needs. That reading decides how deep the answer goes and which method answers it. Here are three real readings:

Examples:

How an answer is put together: behind the answer →

Three rungsthe question decides

Answera fact from your data, back in seconds

Analyticswhy, whether or what next: a tested method from the library

Custom Analyticsa question only your business asks: a full analysis built for your data, yours to keep

you do not have to choosemethod named

THE LADDER · THREE RUNGS

The question decides the rung, not the budget.

Every question about your data is answered on one of three rungs. It starts like a chatbot and does not stop there.

ANSWER When it is a quick question

A total, an average, a count, the highest or lowest group, the value in a given month. It comes back in seconds, read from all of your data, with how it was worked out, and saved so you can find it again instead of asking twice.

ANALYTICS When a question needs more than an answer

What is driving this? Is the difference real? Where is this heading? We run a tested statistical method from our library on your data and interpret it: charts, tables, and a written report that says what the numbers mean and which method produced them. Run it again on next month’s file.

CUSTOM ANALYTICS When it is a custom question

Your own definition of an active customer, your own funnel, the comparison your board asks for every quarter. When no method in the library fits, we build a full analysis for your data, based on proven statistical models. It becomes your own tool, ready the next time the question comes round.

Which rung do I need?

You do not have to choose. Ask the question and we answer at the rung it needs, and we tell you which one that is before any larger piece of work starts. As a guide: a fact from your data is an Answer; why, whether or what next is Analytics; a question specific to how your business works, that you will keep asking, is Custom Analytics. An Answer that turns up something worth pursuing is a good reason to run the Analytics on the same data, and nothing you learned is wasted.

FOR EXAMPLE · THE DOMAIN

Change one word, get a different expert.

Your question carries several signals and each is read on its own. One of them is the field you work in, and you never pick it from a menu, because your own words already say it. Here is the same question three times, changing only the middle:

“How do I measure the incrementality of adding a new line of cars this year?”

a new line of cars

Retail expert. Checks whether the new line is quietly stealing sales from your existing models before it reports a win, and what the same weeks looked like last year.

a second production shift

Manufacturing expert. Checks whether output moved beyond normal process variation, and reads the result against your spec limits.

a new pricing tier

Subscriptions expert. Checks whether the tier grew revenue or just moved existing customers between plans.

The statistics underneath don’t change. The expertise in front of them does, and all you did was describe your own business. If the question isn’t phrased for any one field, that’s a real answer too: you get the generalist, not a guess. The whole story →

ON EVERY RUNG · WHAT COMES BACK

The method is named and the working is shown.

Whichever rung your question lands on, the same four things are true of what comes back.

Simple

You ask in plain words and we pick the method.

Repeatable

The same data gives the same numbers, and you can run it again on new data.

Defendable

The method is named and the working is shown, so when someone asks how you know, you can show them.

Searchable

Everything is saved to your account, and you can ask across all of it later.

WHAT YOU GET · THE DELIVERABLE

It's a link. And it's yours.

Hourly Conversion: Ads vs ControlAnalytics · interactive · open it live

A delivered Analytics report: hourly conversion rate, ad treatment vs PSA control on a public marketing A/B dataset, with a two-line comparison chart, computed rates and a written insight

Calculated, not hallucinated: the numbers were computed before a word was written.

✓Interactive report: explore the charts and diagnostics, not screenshots.
✓The method, stated: what was tested, why that test, whether it held.
✓PDF & one-click citation: APA, MLA, Chicago, BibTeX.
✓The R source: reproducible by anyone, including you.

Open real delivered reports →

AND THEN · IT COMPOUNDS

Every answer puts another landmark on the map.

An analysis doesn’t just answer your question. It finds ideas in your data: that tenure predicts churn, that margin breaks past a discount band, that one cohort behaves unlike the rest. Each becomes a point on a map of your business. The more you ask, the more of the landscape is charted, and the bigger the questions you can ask next.

playing

Your knowledge map1 ANALYSIS · 3 IDEAS CHARTED

One analysis, three ideas. Your first question about discounting didn’t just return a number: it charted the ideas it had to understand to answer it.

This is why the questions get better over time. A brand-new tool can only answer from the data in front of it. A charted business can be asked things that cross several analyses at once, and every answer comes back citing the piece of work that proves it.

And your agents can read the same map. Point whichever AI you use at your library and it navigates by these landmarks, answering from your own saved analyses instead of guessing about your industry. The knowledge layer →

Open source

Read the code that talks to your data.

The MCP server, the tool contracts and the transport options are public under an MIT licence, so you can see exactly what connects to your assistant before you connect it, and run it yourself if you’d rather.

View the repository Read the docs

embeddedlayers / mcp-analytics PUBLIC

The statistical analyst in your AI chat: bring data and a question, own a citable, re-runnable analysis. Works in Claude, Cursor, and any MCP client.

mcp-servermodel-context-protocolclaudedata-sciencestatistics

Python MIT licence 20 topics

WHERE NEXT

proofCase studies →Delivered analyses you can open and read end to end. try itFree analysis widgets →Point one at your CSV and see a result in minutes. price itPricing →What it costs and what it delivers.

Where to start

One email with your data and your question. We take it from there.