Platform overview
GrowthBook is built to be run by agents as well as people. Version 5.1 (September 2026) added a hosted MCP server at mcp.growthbook.io with OAuth sign-in, next to the existing self-hosted option (npm package or Docker image). The server is a thin layer: it loads GrowthBook's open-source agent skills and passes calls through to the REST API, so agents can work across flags, experiments, metrics and dashboards. The same skills install as a plugin for Claude Code, Cursor and Codex. They cover the full flag lifecycle (create, target, ramp, safe rollout with auto-rollback, cleanup) and the experiment lifecycle (brainstorm, design, launch, analyze, stop). Every agent change goes through the same drafts, approvals, permissions and audit logs as a human change. A setup wizard lets a coding agent install the SDK and add the first flag.
In the app, an AI Assistant answers questions about your data, AI Chat in the Product Analytics Explorer builds charts from warehouse data, and a Slack app (beta) lets teams ask @GrowthBook questions and get notifications, such as when a test reaches significance. AI also writes result summaries, SQL and metric descriptions, checks hypotheses against an org-defined prompt, and finds similar past experiments using vector embeddings. Saved Learnings keep conclusions drawn across tests, and an agent can scan past experiments to draft them.
The AI Visual Editor (Chrome side panel, Pro and Enterprise) turns plain-language prompts into test variations, generates images, and can rebuild a Figma frame or mockup as a variation (beta). Contextual bandits (Enterprise, closed beta) learn separate traffic weights for each user segment. Teams can also A/B test their own prompts and models using Langfuse or Arize Phoenix trace data (beta). Admins pick the AI provider and can bring their own key.
.svg)
