GrowthBook MCP Server

The GrowthBook MCP server is the official Model Context Protocol integration from GrowthBook, the open-source feature flagging and experimentation platform. It lets an AI coding assistant (Cursor, Claude Desktop, Claude Code, VS Code, or any MCP-compatible client) read and mutate your GrowthBook configuration directly. You can create feature flags inline as you scaffold a component, wire up an A/B test, generate TypeScript types for your flags, or look up SDK connection details without leaving your editor.

The server exposes tools covering the main GrowthBook primitives: feature flags (create, list, find stale, add force rules, generate types), experiments (list, create, get attributes, manage defaults), environments, projects, SDK connections, metrics, and a documentation search tool. Tool permissions are scoped to the API key or Personal Access Token you provide, so an agent cannot perform actions the underlying user is not allowed to do.

The server is open source, published as @growthbook/mcp on npm and maintained by GrowthBook in the growthbook/growthbook-mcp repo. It runs locally over stdio and can point at either GrowthBook Cloud or a self-hosted GrowthBook instance via GB_API_URL and GB_APP_ORIGIN.

Tools

Prerequisites

• A GrowthBook account (Cloud at app.growthbook.io or self-hosted)
• A GrowthBook API key or Personal Access Token. Capabilities will be scoped to the permissions of that key.
• Node.js installed locally (the server runs via npx)

Required environment variables

• GB_API_KEY: GrowthBook API key or PAT
• GB_EMAIL: Email address tied to your GrowthBook account (used when creating flags and experiments)

Optional environment variables

• GB_API_URL: Defaults to https://api.growthbook.io. Set this for self-hosted installs.
• GB_APP_ORIGIN: Defaults to https://app.growthbook.io. Used to generate UI links.
• GB_HTTP_HEADER_*: Pattern for adding custom HTTP headers, e.g. GB_HTTP_HEADER_X_PROXY=value.

Cursor / Claude Desktop config

{
  "mcpServers": {
    "growthbook": {
      "command": "npx",
      "args": ["-y", "@growthbook/mcp@latest"],
      "env": {
        "GB_API_KEY": "YOUR_API_KEY",
        "GB_API_URL": "https://api.growthbook.io",
        "GB_APP_ORIGIN": "https://app.growthbook.io",
        "GB_EMAIL": "you@example.com"
      }
    }
  }
}

Claude Code (terminal)

claude mcp add growthbook --transport stdio \
  --env GB_API_KEY=YOUR_API_KEY \
  --env GB_API_URL=https://api.growthbook.io \
  --env GB_APP_ORIGIN=https://app.growthbook.io \
  --env GB_EMAIL=you@example.com \
  -- npx -y @growthbook/mcp@latest

VS Code

Use the same JSON as above but under the "servers" key instead of "mcpServers".

• Scaffold a feature flag directly from your editor while writing the code that consumes it, including key, type, default value, and targeting rule.
• Spin up a new A/B test linked to a feature flag, with variations generated by the AI from a short hypothesis.
• Audit and clean up stale feature flags by asking the agent to list flags older than N weeks and propose removals.
• Keep TypeScript flag types in sync with GrowthBook by regenerating types whenever flags change.
• Inspect SDK connections, environments, and metrics to debug why a flag isn't evaluating as expected in a given environment.

• "Create a feature flag called checkout-v2-enabled (boolean, default false) in the web project and wrap the <CheckoutPage /> component with it."
• "List all stale feature flags in production and suggest which ones are safe to remove."
• "Create an A/B experiment for the new pricing page hypothesis, with control and a 50/50 variant, and link it to the pricing-redesign flag."
• "Regenerate the TypeScript types for our GrowthBook flags and write them to src/types/flags.ts."
• "Show me all SDK connections and which environments they target, then create a new Node.js SDK connection for staging."

• Official server maintained by GrowthBook and published as @growthbook/mcp on npm.
• Broad tool coverage spanning flags, experiments, environments, projects, SDK connections, metrics, and docs search.
• Works against both GrowthBook Cloud and self-hosted instances via configurable API and app URLs.
• Permissions are inherited from the API key or PAT, so agent capabilities are bounded by existing user roles.

• Runs over stdio only; there is no hosted remote MCP endpoint, so each user must configure the npx command and env vars locally.
• Tool inputs and outputs are not exhaustively documented in the README; some discovery happens at runtime.
• Capabilities are limited by the underlying API key permissions, which can be confusing when an agent action silently fails due to insufficient scope.

• LaunchDarkly MCP server for teams on LaunchDarkly's feature management platform.
• Statsig MCP server for feature flagging and experimentation on Statsig.
• Using the GrowthBook REST API directly via a generic HTTP MCP server if you need custom tool definitions.