SplitLaunch is a Model Context Protocol (MCP) server and package-first A/B testing platform for customer-owned websites and apps. It lets developers, marketers, and AI coding agents install one package, create URL-based experiments, track conversions, and retrieve performance through an MCP server and agent-friendly command API.
SplitLaunch exposes an MCP server at:
https://www.splitlaunch.dev/api/mcp
Connect via Claude Desktop, Claude Code, or any MCP-compatible client using your SplitLaunch API key as the bearer token. The MCP server provides tools for creating projects, allowlisting domains, launching A/B experiments, tracking conversions, and retrieving performance reports.
The SplitLaunch MCP server exposes tools across these categories:
- projects — create and manage client project workspaces and pixel IDs
- domains — allowlist customer-owned domains for pixel and redirect safety
- experiments — create, update, and manage URL A/B test lifecycle
- goals — create conversion goals for browser pixel or agent firing
- reports — retrieve experiment totals, time-series, and channel breakdowns
- usage — query billing usage and plan limits
- billing — open billing portal for subscription management
The strongest agent resource in this repository is SKILLS.md. It is a readable command playbook for agents using the SplitLaunch MCP server and API. Agents that need the newest command contract should use the live skills URL:
https://www.splitlaunch.dev/api/skills
- Installs into the user's own repository with
@splitlaunch/ab. - Retrieves a public pixel ID through
npx splitlaunch init. - Keeps private API keys in terminal, server, CI, or agent workflows.
- Runs A/B tests between existing control and variation URLs.
- Uses sticky visitor assignment and redirect-based traffic splitting.
- Tracks sessions, exposures, custom events, conversions, revenue, UTMs, click IDs, referrers, device context, browser context, and geo context.
- Lets agents create projects, allowlist domains, create experiments, launch or pause tests, create goals, retrieve reports, and open billing.
SplitLaunch is intentionally small. The dashboard handles account setup, API keys, pixel IDs, projects, usage, billing, and docs. The user's AI agent handles the experimentation work.
This repository is public technical documentation for SplitLaunch. It contains command references, examples, runbooks, and agent workflow guidance.
This repository does not contain the private SplitLaunch application source code. SplitLaunch runs as a hosted platform at splitlaunch.dev.
Install the package in the website or app repository where the SplitLaunch pixel should run:
npm install @splitlaunch/ab
npx splitlaunch initThe CLI asks for a SplitLaunch API key, validates the account, retrieves the
project pixel ID, writes splitlaunch.config.json, and prints the browser pixel
snippet.
For non-interactive agent or CI setup:
SPLITLAUNCH_API_KEY=sl_live_REPLACE_WITH_REAL_KEY npx splitlaunch initFor accounts with multiple projects:
SPLITLAUNCH_API_KEY=sl_live_REPLACE_WITH_REAL_KEY npx splitlaunch init --project PROJECT_IDSee QUICKSTART.md for the short setup path.
API command endpoint:
POST https://www.splitlaunch.dev/api/external/v1/commands
API headers:
Authorization: Bearer sl_live_REPLACE_WITH_REAL_KEY
Content-Type: application/json
Live skills page:
https://www.splitlaunch.dev/api/skills
Hosted docs:
https://www.splitlaunch.dev/docs/install
https://www.splitlaunch.dev/docs/api
Use the canonical www endpoint for scripts. Avoid calling the apex
https://splitlaunch.dev endpoint from API clients because redirects can drop
the Authorization header.
All API-key commands use this envelope:
{
"command": "experiments.create",
"input": {
"projectId": "PROJECT_ID",
"name": "Homepage headline test",
"controlUrl": "https://example.com/",
"variationUrl": "https://example.com/homepage-v2",
"trafficAllocation": 50
}
}High-value command families include:
projects.*for projects, public pixel IDs, allowed domains, and attribution settings.domains.*for allowlisting customer-owned domains before redirect tests.experiments.*for URL A/B test creation, updates, lifecycle changes, and listing.goals.*for conversion goals that browser code can fire.reports.experiment.*for totals, time series, channel breakdowns, lift, and revenue context.usage.summaryandbilling.portalfor plan usage and subscription management.
The skills folder splits the command surface by capability so agents can retrieve exactly the area they need:
Use the runbooks for repeatable workflows:
Use the examples folder for copy-paste starting points:
- Create a project with cURL
- Launch an experiment draft with cURL
- Pull an experiment report with cURL
- Set up tracking with cURL and browser code
- Agent install and launch prompt
- Agent weekly check prompt
The knowledge base gives agents short operational articles:
- Connect an external AI agent
- Set up URL experiments
- Install the pixel and track events
- Read experiment results
Agents should:
- Never put a private API key in browser code.
- Use only the public pixel ID in client-side snippets.
- Resolve the right
projectIdbefore write commands. - Allowlist every control and variation hostname before launch.
- Create experiments in draft first.
- Confirm before setting an experiment to
running, pausing an active test, ending a test, or changing attribution settings. - Read back resources after write commands.
- Explain experiment results with sample size, conversion rate, absolute difference, relative lift, revenue, and caveats.
- Website: https://www.splitlaunch.dev
- Install docs: https://www.splitlaunch.dev/docs/install
- API docs: https://www.splitlaunch.dev/docs/api
- Live skills: https://www.splitlaunch.dev/api/skills