A/B Test Sample Size Calculator MCP

A/B test sample size calculator and significance tests for AI assistants, with exact answers instead of mental arithmetic.

This is a Model Context Protocol (MCP) server. It works out how many users an A/B test needs, how long it must run, whether a result is significant (z-test, chi-square, Fisher exact), the Fisher-Freeman-Halton exact test for larger tables, and how much peeking inflates false positives. It needs no API key and makes no network calls.

Tools

Every result has structured JSON, a one-paragraph summary, as_of, source (the method references) and more (the matching calculator on abwex.com).

What it knows that your model does not

• Exact inverse normal quantiles and the full sample size formula. Models round z-values, drop the pooled variance term or confuse relative and absolute MDE, and the answer moves by thousands of users.
• Why calculators disagree. The result shows both the standard pooled formula (as R power.prop.test) and Evan Miller's published formula, so a 37,513 vs 36,314 gap has an explanation.
• Exact hypergeometric sums. A Fisher-Freeman-Halton p-value for a 4 x 4 table needs millions of tables summed; it cannot be estimated in your head.
• Sequential testing numbers: checking 5 times at p < 0.05 gives a 14.2% false positive rate, and the boundaries that fix it.

Install

Hosted (Claude.ai custom connector, ChatGPT developer mode)

Paste this URL. No token, no sign-up.

https://ab-test-sample-size-calculator-mcp.lipmichal.workers.dev/mcp

Claude Code

claude mcp add --transport http ab-test-calculator https://ab-test-sample-size-calculator-mcp.lipmichal.workers.dev/mcp

Or run it locally over stdio:

claude mcp add ab-test-calculator -- npx -y github:theluckystrike/ab-test-sample-size-calculator-mcp

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "ab-test-calculator": { "command": "npx", "args": ["-y", "github:theluckystrike/ab-test-sample-size-calculator-mcp"] }
  }
}

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "ab-test-calculator": { "url": "https://ab-test-sample-size-calculator-mcp.lipmichal.workers.dev/mcp" }
  }
}

VS Code

Add to .vscode/mcp.json:

{
  "servers": {
    "ab-test-calculator": { "type": "http", "url": "https://ab-test-sample-size-calculator-mcp.lipmichal.workers.dev/mcp" }
  }
}

Node 20 or newer is needed for the local stdio version.

Example prompts

• "Our checkout converts at 4.2%. How many users per variant to detect a 10% relative lift with 80% power?"
• "We get 4,000 visitors a day and want to test 3 variants. How many weeks will it take?"
• "Control had 420 conversions from 10,000, the variant 495 from 10,050. Is that significant?"
• "Run a Fisher-Freeman-Halton test on [[12, 30, 8], [20, 25, 15]]."
• "If I check the dashboard every day for 10 days, what is my real false positive rate?"

Data sources

No live data. Methods and the published values the tests check against are in DATA_SOURCES.md: R power.prop.test and fisher.test documentation outputs, Evan Miller's calculator code, scipy and statsmodels reference values (scripts/reference.py), Armitage et al. (1969) and Jennison and Turnbull (2000).

Limitations

• Sample size covers conversion rates (two proportions). Continuous metrics such as revenue per user need the standard deviation and are not covered.
• Sample size and duration assume a fixed-horizon test analysed once. For several looks, plan with check_peeking_risk.
• The Fisher-Freeman-Halton test enumerates tables exactly. Very large or sparse r x c tables can exceed the work limit (lower on the hosted server than locally). The tool then returns the chi-square result with a note instead of an exact p.
• Peeking boundaries assume equally spaced looks and two-sided tests; boundaries are given for up to 10 looks.
• The confidence interval for the difference is the Wald interval, which is poor for rates near 0 or 1 with small samples.

More

Interactive versions of these calculators and guides are on abwex.com, including the Fisher-Freeman-Halton calculator and the peeking inflation calculator.

Part of the Zovo MCP collection: https://mcp.zovo.one

Development

npm install
npm test                      # builds, then runs the fixture tests
python3 scripts/reference.py  # recomputes the reference values with scipy and statsmodels
node scripts/smoke-stdio.mjs  # stdio check, one call per tool
npx wrangler deploy           # Cloudflare Worker, Streamable HTTP at /mcp

License

MIT

More data servers from the same author

Each one is free, needs no API key and has a hosted URL you can paste into Claude or ChatGPT.

The paid back-office collection (invoices, VAT, PDFs, spreadsheets) is at https://mcp.zovo.one