Bell Statistics
Connect it now

10 free analyses · under a minute

Free public beta

Design and analyze geo tests without leaving Claude

Bell Experiments MCP gives your AI assistant the statistical tools it does not have. Ask it to select treatment and control markets, estimate the required budget, or analyze the results, and it runs the Bell Statistics engine — the same engine behind the Bell Experiments app — instead of estimating a p-value from memory. It works in Claude, Cursor and anything else that speaks the Model Context Protocol today; ChatGPT and A/B testing are next.

  • GeoTests live today
  • A/BComing next
  • 10Free analyses on a trial workspace

Try it now — 10 free analyses

In Claude, open Settings → Connectors, choose Add custom connector, paste this URL and sign in. New accounts get a trial workspace. Cursor, VS Code, Claude Code and the rest are one click away below.

https://mcp.bellstatistics.com/v1
See every way to connect

Built by the statisticians behind experiments at

  • 365Scores
  • Arpeely
  • Flare
  • GlassesUSA
  • Hippo
  • IronSource
  • Lemonade
  • Lumen
  • Lusha
  • Melio
  • Monday.com
  • Moonactive
  • Moonshot
  • MyHeritage
  • Outbrain
  • Papaya
  • Payoneer
  • Playstudios
  • Playtika
  • Plus500
  • ReasonLabs
  • Roundforest
  • Selina
  • Simply
  • Sunflower
  • Webeselense
  • ZipRecruiter
  • ZoomInfo

One test, start to finish

Bringing in the data, designing a geo test, estimating the budget and analyzing the result — in the client you already work in, with every step saved to your Bell Experiments app.

Claude

  1. Here are two years of weekly sales by DMA. Add them to Bell.

    • create_dataset_upload
    • register_dataset

    returns → a dataset in your Bell workspace

  2. Design a geo test for a $200k brand campaign across our US markets.

    • start_allocation_search
    • start_design_analysis
    • get_job

    returns → candidate allocations · pre-period fit

  3. What lift could I detect in these 12 markets on a four-week test?

    • start_design_analysis
    • get_job

    returns → smallest detectable effect · required budget

  4. Read the lift on the test that finished Friday and give me the iROAS.

    • start_results_analysis
    • get_job

    returns → incremental lift · 90% confidence interval · iROAS

    Illustrative: treated markets against their synthetic control. The full analysis opens in Bell Experiments.
    Open Bell Experiments

The prompts and the tool calls are the demo, and the chart is a sketch rather than a result: we do not publish numbers we have not validated, which is the same reason the tools refuse to answer when they cannot. Analyses run in the background while Claude checks on them; if your data is already in Bell, it starts from list_datasets instead of an upload.

Knowing what your marketing actually caused

Marketing teams need to understand incrementality: how much additional business activity was actually caused by a campaign, beyond what would have happened anyway. Geo experiments estimate this by comparing treated and untreated geographic markets. The main challenge is finding a sufficiently balanced allocation that enables high statistical power with a low required budget. Our product helps identify these efficient test designs and measure their incremental impact reliably.

Flawed tracking and biased models make it impossible to know if your marketing campaigns or product changes actually drove real growth. Rather than relying on rigid dashboards or untrustworthy AI guesses, you can now design and analyse statistically sound experiments directly inside your own AI tools. Cut through the noise and measure the true impact of every test without leaving Claude, Cursor or the client you already work in.

How it works

Everything you run in Claude lands in your Bell Experiments app: datasets, designs and results, in the same workspace.

  1. Connect

    Add the server to your AI assistant and sign in. New accounts get a trial workspace with 10 free analyses.

  2. Bring your data

    Your assistant uploads a market-level CSV — daily or weekly, per market — or picks a dataset already in Bell.

  3. Plan

    Select an allocation and estimate the budget — all in plain English.

  4. Analyse

    Measure the exact incremental ROI of your test without ever leaving your chat interface.

What you can ask it

Geo test design and measurement are what the beta runs today. Mid-test health checks and the A/B tools are next.

  • Geo test design

    Find a treatment-control allocation, construct a Synthetic Control, and plan the budget. For questions a user-level split cannot answer — brand, TV, and other market-level campaigns.

    “Design a geo test for a $200k brand campaign across our US markets.”

  • Power analysis

    Minimum detectable effect calculations and budget requirements for regional tests, so you find out what the test can and cannot see before you commit the spend.

    “What lift could I detect in these 12 markets on a four-week test?”

  • Mid-test health checks

    Soon

    Monitor running experiments for unusual activity or data anomalies that could affect the analysis.

    “Here are two weeks of the running test — is anything wrong with it?”

  • Geolift measurement

    Calculating true causal impact and incrementality beyond organic baselines, with the confidence interval alongside the estimate rather than a point figure on its own.

    “Read the lift on the test that finished Friday and give me the iROAS.”

  • Results analysis

    Statistical significance readouts, confidence intervals, and incremental ROAS reporting — written so the uncertainty survives the trip from the analysis to the slide.

    “Write this up for a marketing director, caveats included.”

  • Standard A/B testing

    Soon

    Sample size calculations, SRM detection, t-tests and proportion tests, CUPED variance reduction, and sequential testing. Live as free calculators on this site already; what is pending is exposing them as tools.

    “Size an A/B test at a 3% baseline and a 10% relative lift.”

Connect it in under a minute

One server behind all of them and nothing to install locally. You sign in to Bell Experiments the first time you connect, and a new account gets a trial workspace with 10 free analyses.

  • Claude

    Works in Claude for desktop, web and mobile. Claude Code has a one-line command below.

    1. Open Settings → Connectors.
    2. Choose Add custom connector.
    3. Paste the server URL below and confirm.
    4. Sign in to Bell Experiments when prompted.
    5. Start a new chat and ask it to plan a geo test.
    https://mcp.bellstatistics.com/v1

  • ChatGPT

    Soon

    Not live yet. The steps below are what it will take once it ships — there is nothing to paste until then.

    1. Open Settings → Connectors → Advanced.
    2. Choose Create, and select MCP server.
    3. Paste the server URL we send you.
    4. Enable it in the composer's tools menu.

  • Any MCP client

    Cursor, Zed, Goose, your own agent — anything that speaks the Model Context Protocol over HTTP.

    1. Add the block below to your client's MCP configuration file.
    2. Restart the client so it picks up the new server.
    3. Sign in to Bell Experiments when the client asks.
    mcp.json
    {
  "mcpServers": {
    "bell-experiments": {
      "type": "http",
      "url": "https://mcp.bellstatistics.com/v1"
    }
  }
}

Or do it in one step

Paste a command into your terminal, or click to install in your editor.

Claude Code
claude mcp add --transport http --scope user bell-experiments https://mcp.bellstatistics.com/v1
Codex
codex mcp add bell-experiments --url https://mcp.bellstatistics.com/v1
Gemini CLI
gemini mcp add --transport http bell-experiments https://mcp.bellstatistics.com/v1
Add to CursorAdd to VS Code

Opens the editor, so it needs Cursor or VS Code installed.

Why not just ask the model?

Because a language model will answer a statistics question whether or not it can. It will produce a p-value that looks right, from a test it chose because the question sounded like that kind of question, and it will not mention the assumption the data fails. The failure is silent and it is confident, which is the worst combination available. This server is the part that says no.

It computes, it does not recall
Every number comes from the Bell Statistics engine, called as a tool. Nothing is inferred from the shape of the question, which is what makes the answer reproducible — ask twice, get the same figure.
It refuses to peek
A test read before its planned end does not have a valid p-value, and reporting one anyway is how a team calls a flat quarter a win. It will tell you the test is not readable yet, or design it up front as a sequential test with the correction already applied.
It corrects for multiple comparisons
Five metrics at α = 0.05 is a one-in-four chance of a false positive somewhere. It counts what you are actually testing and applies the correction, rather than treating each question as if it arrived alone.
It tells you when the answer is no
Some questions cannot be answered with the data you have, at the traffic you have, in the time you have. Hearing that in the design phase costs you an afternoon. Hearing it after the test costs you the quarter.

How the geo methodology works

Three stages, in this order. None of it is proprietary in a way that stops us describing it — the work is in doing it well, not in keeping it quiet.

  1. Find an allocation

    A search algorithm explores many possible treatment-control allocations and identifies promising test designs. This is where the budget is won or lost: a well-balanced allocation can deliver statistical power that a weaker design would require more spend to achieve.

  2. Fit the counterfactual

    For each design, the counterfactual outcome of the treatment markets is estimated using a synthetic control — a weighted combination of control markets, with regularisation used to improve stability and predictive performance.

  3. Quantify the uncertainty

    Historical data are repeatedly split and evaluated using out-of-bag sampling, to estimate the experiment's uncertainty, determine the required budget, and quantify the campaign's incremental effect once the test is complete.

One engine, in Claude and in the app

The Bell Statistics engine powers both the Bell Experiments app and this server. Datasets you upload through Claude, and the experiments it designs, open in the app's wizard, in the same workspace — so a test planned in a conversation is already there when you open the app, and nothing has to be rebuilt by hand.

Who it's for

If one of these is your Monday, it is built for you.

  • Growth and marketing analysts

    You are asked whether the campaign worked, and the dashboard answers a different question. This designs the test that answers yours, and tells you what it will cost to be able to see the effect you care about.

  • Data scientists

    You know how to do this and you do not want to hand-roll another market-matching script. The methodology is stated, the numbers come from the same engine as the Bell Experiments app, and the assumptions come back with the answer.

  • Product and marketing managers

    You need the readout to survive the meeting. Ask in plain language, get a design you can defend and a result with the uncertainty still attached to it.

  • Agencies running client tests

    Several clients, several markets, and a measurement story that has to hold up to someone else's analyst. Design and read each test in the assistant you are already writing the deck in.

The questions we usually get

Worth reading before you connect it.

What is an MCP server, and why does statistics need one?
The Model Context Protocol is an open standard that lets an AI assistant call external tools. Statistics needs one because a language model does not calculate — it predicts text. Given a contingency table it produces something shaped like a chi-square result. Given the same table through a tool call, it gets the actual statistic, and the assistant's job drops to the part it is good at: understanding the question and explaining the answer.
Does my data leave my assistant?
The data you analyze reaches our server, because that is what the engine computes on: a market-level time series — daily or weekly values, one series per market. It is stored in your Bell workspace, which is how it shows up in the Bell Experiments app. There is no user-level data in it, the tools are not designed to accept any, and we do not train anything on what you upload. If your data cannot leave your environment at all, email us at info@bellstatistics.com; that constraint is one of the things the beta exists to find out about.
Does it work with my data warehouse?
Not by connecting to it directly — the server has no credentials for your warehouse and is not asking for any. What it takes is the market-level series as a CSV export, which your assistant uploads to your Bell workspace — the data itself is never pasted into the chat. If you already have a warehouse tool connected to Claude, it can run the query, export the result and upload it in the same conversation. That composition is the reason this is an MCP server and not a web app. Direct warehouse connectors are on the list; email us at info@bellstatistics.com with the warehouse you are on, because that is how we decide the order.
Can I self-host it?
Not during the beta. There is one hosted server while we are still learning how the tools behave on real data, and running several versions of it in other people's environments is the fastest way to lose that signal. Self-hosting and a VPC deployment are both on the table afterwards, and demand from beta users is what decides whether they happen early. If your data cannot leave your environment at all, email us at info@bellstatistics.com — you would be telling us something we want to know.
Is it free?
You can start free: signing in creates a trial workspace with 10 free analyses. There will be paid plans for heavier use. We have not set the prices yet, and we would rather say that than publish a number we change.
Which tests and designs does it cover today?
Today: geo testing end to end — test and control market selection, pre-period fitting, detectable-lift estimation at a given spend, and reading the result as incremental lift with a confidence interval and an iROAS. Next: the A/B surface, which is sample size and minimum detectable effect, unequal allocation, sequential designs, sample ratio mismatch, and the two-proportion, two-sample, paired, ANOVA, chi-square, Fisher's, McNemar's, Wilcoxon, correlation, log-rank, equivalence and non-inferiority tests. Those are already live as free calculators on this site; what is pending is exposing them as tools.
Does this replace my experimentation platform?
No, and it is not trying to. Your platform assigns users, holds the flags and collects the events, and it should keep doing that — and for the geo questions the beta runs today, it was never the tool anyway: a market-level holdout is not something a user-level splitter can express. This sits upstream and downstream, deciding what to run before you configure anything and interrogating the result afterwards.
What happens to my existing tests?
Nothing. It does not assign users, hold your flags or collect your events, so there is no migration and nothing to switch off — connecting it changes what you can ask, not what is running. You can point it at a geo test that finished last quarter to re-read the result, and it reads the data you upload to your Bell workspace rather than reaching into anything.
What if it gives me the wrong answer?
Every number comes from the Bell Statistics engine — the same engine the Bell Experiments app runs on — so a computational error is unlikely and, if it happens, is reproducible and fixable. The likelier failure is subtler: the right calculation on the wrong framing of your question. That is what the design-phase checks are for, and it is why the tools return their assumptions alongside the number rather than just the number.
Do I still need statisticians?
For a standard test on clean data, honestly, less often — that is the point of shipping it. What it does not do is design a measurement strategy for a business it has never seen, arbitrate between two teams whose metrics disagree, or build a marketing mix model. When one of those is the actual problem, the tool will say so, and our consulting work is there.
When do A/B testing and the ChatGPT connector land?
Both are built. Neither has been through the validation pass the geo tools have, and we would rather ship one surface we trust than three we are guessing about — the whole argument for this tool is that it refuses to answer when it cannot, which is not a claim you make on untested code. The page changes the day they do rather than before.
How long is the beta, and what do you want from me?
Until the tool surface settles — there is no end date yet. Every new account gets a trial workspace with 10 free analyses, and nothing else is required of you: no scheduled calls and no case-study commitment. If you tell us what you pointed it at and where it disappointed you, at info@bellstatistics.com, that is what decides what we build next.

Rather ask your assistant?

Ask Claude about this pageAsk ChatGPT about this page

Put a statistician in your assistant

Sign in and your trial workspace comes with 10 free analyses, with no obligation afterwards. If you would rather talk to a person about the experiment you are planning, that option has not gone anywhere.

Connect it now