Marketing agents live in ad platforms and analytics. Product experiments — the A/B tests and feature-flag holdouts that decide what actually ships — usually sit in a separate tool nobody thinks to connect. GrowthBook is now a live MCP integration in Cogny, so an agent auditing paid media or reporting on program-level experimentation ROI can pull the product side of the story in the same conversation.

One Secret Key, no OAuth app to register.

TL;DR

GrowthBook is live now. Connect it from your workspace Settings → Cogny Managed MCPs.
11 read-only tools: list_experiments, get_experiment, list_experiment_results, get_experiment_results, list_features, get_feature, list_projects, list_fact_metrics, list_metrics, list_segments, list_data_sources. No write, start, stop or toggle tools are registered.
Works with GrowthBook Cloud (api.growthbook.io/api by default) and self-hosted instances — the API host is configurable per workspace.
Auth is a Secret Key or Personal Access Token (readonly scope is enough) created in GrowthBook Settings → API Keys, sent as a bearer token. No OAuth handshake needed.
list_experiment_results scans the latest computed snapshot for every experiment matching a filter in one call — an agent auditing a whole portfolio doesn't need to fetch each experiment individually.
GrowthBook rate-limits at 60 requests/minute; the proxy surfaces HTTP 429 with a clear message instead of retrying silently.
Available on any paid Cogny plan (Solo $9/mo and above). The agent pays for its own inference; Cogny handles the proxy and credential storage.

What you can do with it

Tool	What Claude can do
`list_experiments` + `get_experiment`	Browse experiments, filter by project, data source or status (`draft`/`running`/`stopped`)
`list_experiment_results`	Latest results snapshot for every experiment matching a filter — scan a portfolio in one call
`get_experiment_results`	Latest computed results for a single experiment, optionally by phase or dimension
`list_features` + `get_feature`	Read feature flags, their rules and rollout values
`list_projects`	Projects used to scope experiments, features and metrics
`list_fact_metrics` + `list_metrics`	Metric definitions, fact-table-based and legacy
`list_segments`	Saved audience segments used in experiment analysis
`list_data_sources`	Connected data warehouses/sources

In concrete terms: "Which experiments are currently running, and which ones already hit statistical significance?" "Is the checkout-redesign feature flag rolled out to 100% yet, or still behind a partial rule?" "Pull the latest results for every experiment in the growth project so I can roll up program ROI." The answers come from GrowthBook's own computed snapshots, not a stale export.

It also feeds the experimentation-roi scheduled template: when a GrowthBook MCP is connected, the template calls it directly for the product/feature-flag leg of the ROI calculation instead of guessing at warehouse tables that usually don't carry holdout data for marketing customers.

How to connect it

In GrowthBook, go to Settings → API Keys and click Create Key.
Choose a readonly Secret Key (or a Personal Access Token) and give it a description, e.g. "Cogny."
Copy the generated key — it is only shown once.
In your Cogny workspace, go to Settings → Cogny Managed MCPs, click Connect on the GrowthBook card, and paste:
- The Secret Key or Personal Access Token
- Your API Host, if self-hosted (GrowthBook Cloud users can leave the default https://api.growthbook.io/api)

The key is stored encrypted, sent only as Authorization: Bearer <key> to your GrowthBook API, and scoped to the single workspace. Disconnect at any time.

On rate limits

GrowthBook's API is rate-limited to 60 requests/minute. That's comfortably within range for interactive agent sessions and scheduled audits, but the proxy surfaces a clear error on HTTP 429 — "Rate limit exceeded" — instead of retrying blindly, so the agent knows to back off rather than loop.

On read-only trust

This integration is strictly read-only, even though GrowthBook's own API supports creating and mutating experiments and feature flags. The proxy registers no create/update/start/stop/toggle tools, so an agent cannot launch an experiment, change a flag's rollout, or touch any GrowthBook state. Every tool carries MCP behavior annotations marking it read-only, so Cogny's UI and any MCP client can see that up front.

Pairing ad-spend data with what's actually shipping in the product is where a lot of ROI conversations get stuck today. Connect GrowthBook from your Cogny dashboard and ask your first question.