GrowthBook MCP: flags and experiments for agents
GrowthBook MCP
Official thin GrowthBook MCP server: the agent reads and changes feature flags and experiments via the REST API and skills
- 24 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- growthbook_api_write runs POST, PUT, PATCH and DELETE and changes flags that affect live users
- The API key opens the whole platform to the agent within the key's rights
Install
- Works in
Manual install
npx @growthbook/mcpLaunch command from the README; set the key through GB_API_KEY in the client config.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- growthbook_api_write runs POST, PUT, PATCH and DELETE and changes flags that affect live users
- The API key opens the whole platform to the agent within the key's rights
This is third-party code. Review the repository files before installing.
What it does
GrowthBook is a feature flag and A/B experiment platform. This server is thin on purpose: it has four tools, while the knowledge lives in a separate GrowthBook skills repository and is bundled at build time. growthbook_list_skills lists top-level skills and growthbook_read_skill returns the wanted workflow, such as feature-flags or its child workflow. growthbook_api_read sends GET requests to the GrowthBook REST API and is marked read-only. growthbook_api_write runs POST, PUT, PATCH and DELETE and is marked destructive so clients can ask for confirmation. The server starts over stdio with npx @growthbook/mcp or as an HTTP server with OAuth. For Cursor there is a marketplace plugin that connects to the hosted server at https://mcp.growthbook.io/mcp and opens a GrowthBook sign-in. For a self-hosted GrowthBook, set GB_API_URL.
Who it is for. Teams that run flags and experiments in GrowthBook and want to change them from Claude Code or Cursor without extra tabs.
Good fit when
- A new feature flag is created step by step from a skill and the agent should find the right workflow at once
- You want to see the list of experiments and their status without opening the dashboard
- The same flag is turned on in a test environment and its targeting rules are checked
- On a self-hosted GrowthBook the agent is given a key and the address of your own API
- A service client should see only the API, so skills are turned off with GB_SKILLS_ENABLED
Not a fit when
- You want a separate tool for each API method: the server offers only universal read and write by path
- You need a guarantee that a flag will not touch live users: the decision stays with you and writes run on the agent's request
- You want to call the gb-call command: the server does not run it and maps GET and POST to its own tools
Example request
Create a new-checkout flag for the test environment enabled for ten percent of users and show the resultMCP
- Transport
- stdio, http
- Authentication
- API key
Environment variables
GB_API_KEYrequired, secret- GrowthBook API key or personal access token, required for stdio.
GB_API_URL- API address, https://api.growthbook.io by default; for self-hosted installs.
Limitations
Stdio mode needs GB_API_KEY, a GrowthBook API key or personal access token. HTTP mode with OAuth makes the key optional but requires GB_MCP_URL, otherwise the server will not start. The API address defaults to https://api.growthbook.io; self-hosted installs set GB_API_URL.
Every REST call carries headers with the tool name, version, transport and client: GrowthBook records them in its usual telemetry, while the server itself sends nothing anywhere. Skills are bundled at build time from the growthbook/skills repository, and scripts inside skill folders are not copied.
How to disable. Remove the growthbook entry from the client config or the plugin in Cursor and revoke the API key in GrowthBook. On a self-hosted GrowthBook, DISABLE_TELEMETRY on the backend turns telemetry off.
FAQ
How do I connect GrowthBook MCP?
Run npx @growthbook/mcp with GB_API_KEY in the client config, or install the GrowthBook plugin from the Cursor marketplace. The plugin uses the hosted server https://mcp.growthbook.io/mcp.
Why does the server carry skills?
Skills hold workflows and guardrails, such as creating a flag. The agent reads them with growthbook_read_skill and then calls the REST API. If you do not need them, GB_SKILLS_ENABLED=false leaves only growthbook_api_read and growthbook_api_write.
Does the server work with self-hosted GrowthBook?
Yes, set GB_API_URL to your API address. For the Cursor plugin a self-hosted setup is configured by hand.
How do the two paths differ in HTTP mode?
The /mcp path gives skills and the API, while /mcp/api gives only growthbook_api_read and growthbook_api_write.
Does the server send telemetry?
Not itself. Every REST request carries X-GB-MCP-Tool, Version, Transport and Client headers, and GrowthBook records them in its own telemetry.
Discussion
No questions or reviews yet.
Related
MCP serversHumblytics MCP gives an agent live site analytics, funnels and A/B tests plus 12 marketing skills for finding conversion problems.
AsoBeast MCP server: an agent reads App Store and Google Play keyword ranks, competitors and the ASO queue of your instance.
One MCP server for ad accounts, analytics and CRM: Google Ads, Meta Ads, GA4, TikTok Ads and LinkedIn Ads
The Betterlytics web analytics MCP server: an agent reads traffic, funnels, JavaScript errors and uptime monitor state.
MCP server for Google Ads: analyze campaigns, metrics and keywords with GAQL queries right in the chat with an agent
HitKeep is open source cookieless web analytics in one binary with a read-only MCP server that gives assistants aggregate data.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| GrowthBook MCP: flags and experiments for agents (this card) | yes | API key | 24 | |
| LaunchDarkly MCP Server | yes | API key | 27 | a server for LaunchDarkly flags, whereas GrowthBook has its own REST API and an open-source platform |
| Flags SDK | yes | 629 | a toolkit for flags in Next.js application code, with no management of an experimentation platform |
README in short
The README covers the four tools, the Cursor plugin, running and environment variables, HTTP mode with OAuth, API-only mode, skill bundling and usage headers. The license is MIT.