Skip to main content

GrowthBook Agent Skills

Agent Skills are folders of instructions, scripts, and resources that an AI agent loads on demand to carry out a specialized task. Anthropic introduced the format, and it works in any agent that follows the Agent Skills standard, including Claude Code, Cursor, and Codex.

GrowthBook publishes a library of them that runs the entire feature flag and experiment lifecycle from your AI coding agent (create a flag, roll it out, run an experiment, read the results, and clean up afterward) without leaving your editor and without running an MCP server.

Each skill is a playbook your agent follows: it knows the GrowthBook REST API, the right order of operations, and the guardrails (draft → review → publish, safe rollouts, two-step deletes) that keep changes safe.

Open source

The skills live at github.com/growthbook/skills and call the GrowthBook REST API directly through a small bundled helper.

Watch GrowthBook's Head of Experimentation, Luke Sonnet, run an experiment end-to-end using skills.

Prerequisites​

  • A GrowthBook account: GrowthBook Cloud or a self-hosted instance.
  • A Personal Access Token (PAT): create one at Account → Personal Access Tokens. The token is tied to your GrowthBook user, so flags and experiments the skills create are attributed to you automatically.
  • Node.js 18+: check with node -v. Many agents already run on it.

Installation​

1. Install the plugin​

Add the marketplace, then install the plugin:

/plugin marketplace add growthbook/skills
/plugin install growthbook@growthbook-skills

2. Configure your credentials​

Run the setup skill and follow the prompts:

/growthbook:gb-setup

It asks for your PAT or API key (and, for self-hosted, your API URL), validates them against the live API, and writes ~/.config/growthbook/.env with chmod 600. Every other skill reads that file automatically.

Prefer environment variables?

The skills read environment variables first and only fall back to the file. Export them instead if you like. Handy for CI:

export GB_API_KEY=<your-key>             # required: PAT or Secret Key
export GB_API_URL=https://api.your-host  # self-hosted only

3. Verify​

/growthbook:flag-search

This should list your existing feature flags. If something's off with the credentials, the error points you back to /growthbook:gb-setup.

How skills work​

Skills fire two ways:

  • Automatically, when your agent detects matching intent: "create a feature flag for the pricing page" runs flag-create; "what should we test next?" runs experiment-brainstorm.
  • Explicitly, by typing the slash command, e.g. /growthbook:flag-search or /growthbook:experiment-launch.

Because each skill names its trigger phrases and routes to sibling skills, they compose cleanly when chained:

  • Flag-first: flag-create → flag-toggle → flag-targeting → flag-ramp / flag-monitoring → flag-cleanup
  • Experiment-first: experiment-design → experiment-launch → experiment-analyze → experiment-stop → flag-cleanup
  • Experiment on an existing flag: flag-experiment → experiment-launch → experiment-stop → flag-cleanup

What's included​

The skills fall into three groups. The two reference pages list each one with an example prompt you can run.

  • Setup — gb-setup configures and validates your API credentials.
  • Feature flag skills — create, target, roll out, ramp, monitor, review, publish, search, and clean up flags across their full lifecycle.
  • Experiment skills — brainstorm, design, launch, analyze, and stop A/B tests.

Example prompts​

  • "Create a boolean flag new-checkout-flow that defaults to false, then roll it out to 10% of US users."
  • "What feature flags are stale and safe to clean up?"
  • "Design an A/B test for the new pricing page and launch it on the pricing-v2 flag."
  • "What are the results of our checkout experiment? Should we ship the winner?"
  • "Stop the homepage-hero experiment and roll the winning variation out to everyone."

Safety and governance​

The skills follow GrowthBook's built-in safeguards:

  • Draft → review → publish. Every flag change is staged on a draft revision. Skills route through flag-review and flag-publish so approvals and merge conflicts are handled the same way they are in the app. See Publishing & approval flows.
  • Two-step deletes. flag-cleanup archives a flag, pauses for you to verify nothing broke, and only then deletes. It won't collapse those steps.
  • Read-only by default where it matters. flag-search, flag-graph, experiment-brainstorm, experiment-design, and experiment-analyze never write.
  • Secrets stay local. gb-setup writes your key to a 0600 file and recommends a scoped PAT you can revoke independently.

Next steps​

Join our community Slack for additional help, tips, and announcements