Launchdarkly Experiment Setup

You’re using a skill that guides you through setting up and running experiments in LaunchDarkly. Your job is to design the experiment, create it with the right metrics, treatments, and flag config, start data collection, evolve the design between iterations when needed, and stop with a winner.

Prerequisites

This skill requires the remotely hosted LaunchDarkly MCP server to be configured in your environment.

Required MCP tools:

• create-experiment — create a new experiment with its initial iteration (hypothesis, metrics, treatments, flag config).
• start-experiment-iteration — begin collecting data for an experiment’s current draft iteration.
• get-experiment — check experiment status, treatments, metrics, and current iteration.

Optional MCP tools:

• list-experiments — browse existing experiments in the project.
• update-experiment — update fields on the experiment or its current iteration. Honours mutableFieldsByStatus, so what’s editable depends on whether the iteration is not_started, running, or stopped. Returns rejected inputs under skipped.
• save-and-start-experiment-iteration — the API-recommended way to change locked fields on a running experiment. Stops the current iteration, creates a new draft with the supplied field updates, and starts it in one call.
• stop-experiment-iteration — stop the running iteration. You must declare a winner: pass the winningTreatmentId (and a winningReason). If no variation outperformed, pick the baseline/control as the winner.
• list-metrics, create-metric, list-metric-events — manage metrics referenced by the experiment.

Core Concepts

What Are Experiments?

Experiments in LaunchDarkly measure the impact of feature flag variations on key metrics. An experiment consists of:

• Treatments: the flag variations being compared (control vs. test). Each treatment has an allocationPercent; the values across treatments should sum to 100.
• Metrics: what you’re measuring (conversion rate, latency, revenue, etc.). One must be the primary metric.
• Flag config: the flagKey, ruleId, and flagConfigVersion of the targeting rule that drives the experiment.
• Iteration: a single data-collection window. Created in not_started status, becomes running when started, transitions to stopped when ended.
• Holdout (optional): a project-level group of users excluded from the experiment for baseline measurement (holdoutId).

Experiment Lifecycle

1. Create the experiment with its first iteration (create-experiment).
2. Start the iteration to begin data collection (start-experiment-iteration).
3. Monitor results as data accumulates (get-experiment).
4. Evolve the design mid-experiment if needed — change locked fields like treatments, metrics, or methodology by calling save-and-start-experiment-iteration, which stops the current iteration, creates a new draft with your changes, and starts it.
5. Stop the iteration when you have a winner or a clear call (stop-experiment-iteration).
6. Ship the winning variation.

Core Principles

1. Metrics first: ensure the metrics you’ll reference exist before creating the experiment.
2. Clear hypothesis: every iteration requires a hypothesis string; state what you expect to improve and by how much.
3. Proper controls: exactly one treatment must have baseline: true.
4. Sufficient sample size: let iterations run long enough for statistical significance.
5. One change at a time: test one variable per experiment for clear attribution.

Workflow

Step 1: Prepare Metrics

1. Use list-metrics to find existing metrics.
2. If you need a new one, use create-metric and note the key.
3. Decide which is the primary metric (a single metric or a funnel group). You’ll pass its key as primarySingleMetricKey or primaryFunnelKey on the iteration.

Step 2: Identify the Targeting Rule

You need the ruleId and current flagConfigVersion of the flag rule that will drive the experiment. Use get-flag on the flag (or its environment-scoped status) to find them. The fallthrough rule’s id is the string "fallthrough".

Step 3: Create the Experiment

Call create-experiment. The top-level fields describe the experiment; the nested iteration object describes the first data-collection window.

{
  "projectKey": "my-project",
  "environmentKey": "production",
  "key": "checkout-flow-v2-experiment",
  "name": "Checkout Flow v2 Experiment",
  "description": "Compare the redesigned checkout against the current flow.",

Useful optional top-level fields:

• holdoutId — attach an existing holdout.
• dataSource — "launchdarkly" (default), "snowflake", or "databricks".
• methodology — "bayesian" (default), "frequentist", or "export_only".
• analysisConfig — set thresholds, multiple-comparison correction, or sequential testing.

Useful optional iteration fields:

• attributes — array of context attribute keys to slice results by (e.g. ["country", "device"]).
• covariateId — covariate CSV id for stratified sampling.
• canReshuffleTraffic — defaults to true; set false to lock users to their initial variation when allocations change.

Step 4: Start Data Collection

{
  "projectKey": "my-project",
  "environmentKey": "production",
  "experimentKey": "checkout-flow-v2-experiment"
}

Before starting, the API requires that:

• the flag is toggled on,
• the iteration has a randomizationUnit, and
• at least one treatment has a non-zero allocationPercent.

Pass changeJustification if you’re restarting after a prior iteration was stopped.

Step 5: Verify

1. Call get-experiment and confirm currentIteration.status === "running".
2. Check that treatments are present with the expected allocations.
3. Check the metric list and the primary metric.

Step 6: Evolve the Design Mid-Experiment (when needed)

Most structural fields (treatments, metrics, methodology, hypothesis, …) are locked while an iteration is running. Two ways to change them:

• Light edits while running — update-experiment will let through anything mutableFieldsByStatus permits in the running state (typically just metadata like name, description, maintainerId, tags, plus appending metrics/attributes). It surfaces rejected fields under skipped with a reason.
• Real design changes — call save-and-start-experiment-iteration. It stops the current iteration, creates a new draft with the supplied field updates applied, and starts it in one call. Inputs match update-experiment, plus changeJustification. Mutability is checked against not_started since updates land on the new draft.

Example: swap the treatment allocation and add a metric in a single call.

{
  "projectKey": "my-project",
  "environmentKey": "production",
  "experimentKey": "checkout-flow-v2-experiment",
  "changeJustification": "Lowering control allocation now that variant looks safe.",
  "treatments": [
    {
      "name": "Control",
      "baseline": 

Step 7: Stop the Iteration

When you’ve reached significance or made a call, stop the iteration. A winning treatment is required to stop — LaunchDarkly does not let you end an iteration without declaring a winner. Pass the winning treatment’s id (returned in get-experiment as _id on each treatment) plus a winningReason.

If the experiment was inconclusive or no variation beat the control, declare the baseline/control treatment as the winner and say so in winningReason (e.g. “Inconclusive — no significant lift, keeping control”). There is no “stop without a winner” path.

{
  "projectKey": "my-project",
  "environmentKey": "production",
  "experimentKey": "checkout-flow-v2-experiment",
  "winningTreatmentId": "treat-002",
  "winningReason": "Two weeks of data, +4.1% lift on the primary metric with PBBL > 95%."
}

Report results:

• Iteration stopped with the declared winningTreatmentId (the control/baseline if inconclusive).
• Lift / significance summary on the primary metric.
• Next steps (ship the winner, roll back, or start a follow-up iteration).

Edge Cases

What NOT to Do

• Don’t omit iteration on create-experiment — it’s required.
• Don’t set baseline: true on more than one treatment.
• Don’t let allocationPercent values fail to sum to 100 across treatments.
• Don’t try to change locked iteration fields with update-experiment while the iteration is running — reach for save-and-start-experiment-iteration instead.
• Don’t stop iterations early — wait for statistical significance.
• Don’t run multiple experiments on the same flag at the same time without a careful holdout design.