FOR_PRODUCT_TEAMS // EVIDENCE OVER OPINION

Stop shipping on opinion.
Run experiments your team can defend.

An experiment OS for product teams. Define the hypothesis, attach the metric, set a kill threshold, then ship a verdict, not a debate.

Run your first product experimentSee it inin 30 seconds
MCP for your agents · unlimited archive · no card required
THE_RULES // four constraints, in order
01Hypothesis
Make a clear hypothesiswhat you expect, why
02Metrics
Attach real metricsnumbers, not vibes
03Execution
No mid-way changesprotect the test
04Verdict
Kill or scaledecide, don't drift
WHAT_CHANGES // shape of the work

From drift to discipline.

The work doesn’t get harder. It gets explicit.

Instead of
try something→wait→feels unclear→switch
You get
define→run→evaluate→decide
explicit · 01You know what you’re testing.
explicit · 02You know when it ends.
explicit · 03You know what the outcome means.
START_WITH_AN_IDEA // ideas are not the problem

Most marketing starts like this.

A passing thought, a screenshot from someone else’s launch, a hunch in the shower. None of them is wrong. None of them is enough.

The problem is not ideas. The problem is what happens next: they stay vague, with no structure, and no clear way to evaluate.

01“Maybe X will work”
02“We should try content / ads / X”
03“I saw someone do this”

What’s missing isn’t the idea, it’s the contract behind it.

  • no hypothesis
  • no metric
  • no kill threshold
  • no end date
AGENT_INTEGRATION // MCP CONNECTOR

Let your agent run it. Keep the discipline.

Xi exposes a remote MCP server. Your agent connects once and drives the same lifecycle you would: commit a contract, log values on cadence, confirm ship or kill, capture the learning.

OAuth via Clerk. 30 seconds to connect. Tokens are scoped to your workspace; nothing leaks past the experiments you authorize.

Same rigor whether a human commits the experiment, an agent does it for you, or a cron job lets the contract decide. One discipline.

Discipline doesn’t need a dashboard.
MCP CONNECTOR · READYxi_mcp/v1
Add xi to your client
https://xiexperiments.com/api/mcp/mcp
Add a custom connector in ClaudeOpen Claude
Open Claude, add the MCP URL, and sign in once.
Try once connected
  • Help me commit my first experiment.
  • Suggest 3 marketing experiments I should run this month with kill thresholds.
You’ll be asked to sign in or create a free xi account during setup.
Using another MCP client?

Use the same MCP URL in your client’s MCP or connector settings.

Older clients without remote-connector support: bridge via npx mcp-remote https://xiexperiments.com/api/mcp/mcp.

HOW_IT_WORKS // decide once, at the start

You decide once. After that, the rules don’t move.

Three steps. The shape doesn’t change between a marketing test, a pricing change, an outreach push, or a habit. The contract is the same.

① Define

Decide what success and failure mean.Before you run anything, write down the contract.

You commit, up front, to what a win looks like, what kills the idea, and how long you’ll run it. Once it starts, you don’t change the rules when the numbers come in.

Takes a few minutes. Templates for common cases: onboarding, pricing, outreach, content, habits.
② Track

See everything in one place.Metrics, notes, and the contract live together.

Your agent can monitor it through MCP: check thresholds as data updates, flag when a kill condition is hit, and log the right data for you.

No constant checking. The system speaks up when the contract says it should.
③ Decide

Let the rules make the call.Time’s up, or a threshold is hit.

The outcome is already defined: ship, kill, or review. Each result is saved with the evidence behind it. Wins and failures both become something you can reuse.

Exports (coming soon): Markdown, JSON, HTML logs.
EXAMPLES // experiments compound
exp_dm_02on track

LinkedIn cold DM · 1 line vs. 3 paragraphs

If I cut my LinkedIn cold DM from 3 paragraphs to a single line, then reply rate will rise from a 7% control to 15% or higher within 40 sends.

reply_rate (%)18/15
count · ↑ higher
6d in4d left
exp_lp_07on track

Landing page · 5 features vs. 1 clear promise

If the landing page leads with one clear promise instead of five features, then signup rate will rise from a 10% control to 12% or higher across 1,000 visits.

signup_rate (%)14/12
count · ↑ higher
4d in6d left
exp_tw_11learned

Twitter growth · daily posting vs. 2 high-quality threads

If I post 2 high-quality threads weekly instead of posting daily, then engagement rate will rise from a 3.9% control to 5% or higher per post.

engagement (%)6.2/5
count · ↑ higher
6d in0d left

Take one idea. Turn it into an experiment.

See what actually happens.

Run your first product experimentSee it in
MCP for your agents · unlimited archive · no card required
// FURTHER_READING
// COMPARE

Xi vs Statsig for marketing-side experiments

Where Statsig wins on product flags and stats, and where Xi wins on the contract.

// FIELD GUIDE

Growth loops, NSM, and the experiment portfolio

The marketing layer for PMs: NSM, growth loops, and how to run an experiment portfolio.