Masset MCP Tools

Free MCP tools for marketers, from Masset. Each tool renders an interactive card inside your AI chat using MCP Apps, the interactive-UI extension that became official in the 2026-07-28 MCP spec.

Nothing you send is stored. Every tool is stateless, deterministic, and open source.

Did It Win? — honest A/B test verdicts

You tell your AI the numbers from your A/B test. It answers with a visual verdict card: WINNER, LOSER, NOT YET, or NO REAL DIFFERENCE, with real statistics computed in code (never estimated by the model).

The card shows:

• Both conversion rates with their uncertainty ranges, so you can see whether they overlap

• The probability the variant is genuinely better (the sentence marketers actually want)

• How many more visitors the test needs before you can trust it, and roughly how many days that is

• A confidence slider you can drag to feel what "95% confidence" trades away

• A warning against peeking, because calling tests early is the #1 way teams ship false winners

Try it in one minute (claude.ai or Claude Desktop)

1. Open Settings → Connectors → Add custom connector

2. Paste: https://mcp.getmasset.com/did-it-win/mcp

3. Ask: "Control got 168 signups from 4,210 visitors. Variant got 203 from 4,305. Did it win?"

Claude Code

claude mcp add did-it-win --transport http https://mcp.getmasset.com/did-it-win/mcp

Tools

The math

All deterministic, all in tools/did-it-win/stats.ts, all unit-tested against scipy/statsmodels reference values (see stats.test.ts and scripts/gen-reference.py):

• Pooled two-proportion z-test (two-sided) — the significance gate

• Wilson score intervals — the uncertainty bars

• Bayesian probability-to-beat — exact closed form with flat Beta(1,1) priors (normal approximation above 20,000 conversions)

• Classic two-proportion sample-size formula — the finish line (80% power by default)

Honesty notes: the "visitors needed" figure is the sample required to detect the observed difference, so it moves as your data moves. When detecting the observed gap would take over a million visitors per arm, the tool says "no real difference" instead of stringing you along. And the verdict gate is fixed-horizon statistics: decide your end date up front, then check.

Related MCP server: Kumiki A/B

Self-host

git clone https://github.com/BenMasset/masset-mcp-tools
cd masset-mcp-tools
npm install
npm start          # builds the card + serves http://localhost:3006/did-it-win/mcp
node smoke.mjs     # end-to-end check against the running server

npm run serve -- --stdio runs the Did It Win server over stdio for local clients.

Example numbers in this README are fictional.

Roadmap (v2 candidates)

• A/B/n (multiple variants, with correction)

• Revenue-per-visitor tests (continuous outcomes)

• Sequential testing support, so peeking is managed instead of just warned about

License

MIT. Built by Masset, the home for your business content.