Flags and Experiments together
Flags control what ships. Experiments measure what changes. splitch keeps both under the App and Environment model so a rollout and an Experiment Run do not drift apart.
Feature flags and A/B experimentation, agents first
Ship the split. Measure the truth.
splitch gives agents and humans one control plane for Flags, Experiments, Environments, and Metrics, with rigor built into the workflow instead of bolted on after launch.
agents connect at mcp.splitch.dev · humans sign in at app.splitch.dev
flag.checkout-ctaRun #7 · live
Control
50%
6,241 exposures
Treatment
50%
6,208 exposures
Flags + Experiments
One product surface for what ships and what proves it.
Agent-first operation
The remote MCP server is the primary agent door, with the CLI and panel as first-class skins over the same typed control-plane contract.
Edge-shaped serving
Evaluation, ingest, analysis, and control-plane Workers stay separate so the serving path can stay fast while authoring remains auditable.
Statistical rigor
Results have to trace back to facts.
splitch treats analysis as an enforced product contract. A result should be auditable back to raw Exposure events, the Experiment Run, and the Metric definition that produced it.
- Sequential, always-valid inference by default
- Exposure is the analysis denominator
- SRM, CUPED, winsorization, and FDR are product contracts
- Bad config fails loud instead of falling back silently
checkout-conversion · lift95% CI
0%+4.2%+8%
- interval
- [1.1, 7.3]
- p-value
- 0.003
- SRM
- healthy
Quickstart
The first run ends on verify, not hope.
Four steps from zero to a measured result. The same quickstart is exposed to agents through the MCP resource splitch://quickstart.
Step 1
Authenticate
Agents connect through the remote MCP OAuth handshake. Humans can use the CLI or panel.
Step 2
Create an App
dev and prod Environments are provisioned for the common first-run path.
Step 3
Create a Flag
Define Variants, select dev, and verify one Targeting Key before wiring code.
Step 4
Measure
Start an Experiment Run and read results against Exposures, not guesses.
Give agents the same flag and experiment surface humans get.
Start in the panel, then let agents operate through MCP with typed tools and the same authority model.