Growth Method
Get startedSearch
Documentation menu

Connect an integration

Integrations give the AI new tools to work with, such as your analytics, your CMS, or your site. This guide shows you how to connect one, then how to choose what it can do and understand what it exposes: its tools, safety hints, server instructions, external websites, and source.

Before you start

  • You’re a manager of your team. Anyone on the team can open Add new integration to see what’s available and what’s already connected, but only a manager can connect one.
  • If the integration uses an API key or other credentials, have them from the provider ready.

Connect an integration

  1. Go to Settings, then Integrations.
  2. Select Add new integration. The panel has its own address, app.growthmethod.com/settings/integrations?add=true, so you can send a colleague straight to it.
  3. Choose the integration you want from the list. Ones your team already has are marked Added.
  4. Finish connecting. What you do here depends on the integration:
    • Sign in with the provider. The integration opens the provider’s sign-in page. Sign in and approve access, and you return to Growth Method, connected.
    • Enter an API key. The integration shows its server details and asks for your API key. Paste it, then select Connect.
    • Enter your credentials. A few integrations ask for more than one value, such as Marketo Engage’s Client ID, Client Secret and Munchkin Account ID. Enter each one, then select Connect.
    • No credentials needed. Some integrations connect as soon as you select Connect.
  5. If the integration offers a choice of what to use, such as which analytics property, you’re asked to select it. Pick what you want the AI to use, then finish. Until you do, the AI can’t use the integration, and afterwards it only reaches what you picked.

Note: Bing Webmaster Tools uses an API key rather than a sign-in. In Bing Webmaster Tools, select Settings, then API access, accept the terms if asked, and select Generate API key. One key covers every site you have verified there, so after you paste it Growth Method asks which site to report on.

Note: Google BigQuery connects with a Google sign-in and reaches every project that Google account can. Its tools work on one project at a time and the AI has no way to list your projects, so name the project ID when you ask, for example “in project acme-analytics”. Queries run on your BigQuery account, so Google’s usual charges for the data they scan apply. Listing datasets and tables and running read-only SQL happen without approval; a query that changes data, or cancelling a job, asks for your approval in chat.

Note: Brevo uses an MCP token rather than a sign-in. In Brevo, go to Account, then SMTP & API, then API Keys, and generate a key with the MCP option ticked. The integration covers your contacts, lists and segments, email, SMS and WhatsApp campaigns, templates, senders and domains, and CRM deals, companies and tasks. Anything that sends or changes something in Brevo asks for your approval in chat.

Note: GrowthBook Cloud is GrowthBook’s hosted service. If you run GrowthBook yourself, run GrowthBook’s official MCP server against your installation with OAuth switched on, then add it with Add custom integration and Sign in (OAuth).

Note: MKT1 signs you in with the email of your MKT1 newsletter subscription, so use that address exactly. Your subscription level is checked when you connect: free subscribers get newsletter, template and job board search plus a few research skills, and paid subscribers get every skill. If you upgrade later, open Edit on the integration and select sign in again to pick up the new level.

Note: Mailchimp Transactional uses a Transactional API key, the one for transactional email (formerly Mandrill), not a Mailchimp Marketing key. If you restrict the key to certain API calls, also enable its AI Agents permission group, or the connection fails. The integration covers transactional sending only: stats, message search, failed sends and templates, not Marketing audiences or campaigns.

Note: Marketo Engage uses the Client ID and Client Secret of a LaunchPoint API service. In Marketo, go to Admin, then LaunchPoint, and select View Details on the service. Adobe only accepts connections from Marketo instances it has approved, so if connecting fails, ask your Marketo MCP administrator to enable MCP access for your Munchkin Account ID.

Note: some providers need an administrator to switch on MCP access before anyone can sign in. In Braze, an admin enables MCP OAuth access under Admin Settings, then OAuth, and grants you the Use MCP Server permission. In Outreach, an admin switches on MCP Server under Administration, Organization, then Org Info, and your organisation needs the Amplify add-on. Klaviyo needs the Owner, Admin or Manager role, and 6sense needs the Revenue Marketing platform. Leadfeeder needs the Manage API keys and use API access right under Integrations, which comes with API access on paid plans. For WordPress.com, switch on Enable MCP access at wordpress.com/me/mcp, where you also choose which write tools to allow. Atlassian lets an organisation admin limit which apps can use its MCP server, so if Jira or Confluence sign-in is refused, ask them to allow Growth Method in Atlassian Administration. If sign-in fails with one of these, check that first.

To connect a remote MCP server that isn’t in the list, select Add custom integration at the top of the list and enter its name and server URL. Then choose how it connects: Sign in (OAuth) opens a browser sign-in with the server, or API key lets you paste a key the server gave you. If the server is public and needs no key, choose API key and leave the field blank. With sign-in there’s no key to copy, and Growth Method keeps the connection fresh for you.

Note: HubSpot asks for the details of an app you create in your own HubSpot account. In HubSpot, go to Development, then MCP Auth Apps, and select Create MCP auth app. Give it a name and, for the redirect URL, paste the address shown in the Growth Method connect form. Copy the app’s Client ID and Client secret into Growth Method, then select Continue to sign in. The admin of your HubSpot account needs to connect before other users in the account can.

Note: WordPress (self-hosted) connects to your own site through the WordPress MCP Adapter plugin. Your site needs WordPress 6.9 or later, the MCP Adapter plugin, and the WordPress AI plugin with Custom Abilities switched on under Settings, then AI. Create an Application Password under Users, Profile, then Application Passwords, and enter your site URL, your username and that password. The site must use HTTPS. The tools you get are the abilities your site has registered: the AI plugin adds read-only content, settings and users, and other plugins can add more, including ones that change your site, which ask for approval in chat.

Note: BetterDocs connects to the MCP server built into the BetterDocs plugin on your own WordPress site. In WordPress, go to BetterDocs, then MCP, and switch on Enable MCP access. In Growth Method, enter your site URL, which must use HTTPS, then select Continue to sign in. A WordPress administrator approves the connection, so there’s no password to copy. Anything that creates, changes or deletes docs or FAQs asks for your approval in chat.

Note: Odoo connects to the MCP server built into your own Odoo database. First, an Odoo administrator goes to Settings, then Integrations, then MCP OAuth, and switches on Allow Dynamic Client Registration. In Growth Method, enter your database URL, such as https://yourcompany.odoo.com, then select Continue to sign in and sign in as an internal user. Out of the box the tools only read your records. An administrator can offer more by ticking Available in MCP on a server action, and any that change records ask for your approval in chat.

Note: when you sign in to Stripe, it asks which accounts or sandboxes to share and what each may do. Grant read-only access unless you want agents to make changes in Stripe. Anything that changes something asks for your approval in chat, and Stripe also holds refunds and money movement until you approve them in the Stripe Dashboard. A Stripe administrator can switch MCP access off for the whole team in Stripe’s Dashboard settings.

Note: a few providers only let apps they already know sign in. If Sign in (OAuth) fails saying a redirect or address isn’t allowed, the provider hasn’t opened its sign-in to Growth Method yet. Connect with an API key instead, or ask the provider to allow Growth Method.

Once connected, the integration appears in your list marked Connected. Select Edit on it, or View on the built-in one, to manage what it can do. The rest of this guide explains what you’ll find there.

Note: to change what an integration is scoped to later, such as which analytics property it uses, open Edit on it and select sign in again. A fresh sign-in always starts from scratch, so you pick the property or site again. This is deliberate, so a connected integration’s access can’t be quietly widened.

Tools found on your site

When your workspace was prepared, the AI read your website and noted the tools it found there, such as your analytics, tag manager, CRM or site platform. They appear in your integrations list below the ones you have connected, each marked Found on your site, so your first connections are a list to work through rather than a catalogue to search.

  1. Go to Settings, then Integrations.
  2. Find a tool marked Found on your site and select Connect.
  3. Finish connecting as above: sign in with the provider, or enter your API key.

Once connected, a tool moves up to join your other connected integrations. Anyone on the team can see the found tools; only a manager sees Connect and Dismiss. If a tool was found that you do not use, select Dismiss and it leaves the list for the whole workspace. Nothing is created until you connect.

What an integration can see

An integration sees whatever the account you connect it with can see. If that account reaches several sites, projects or clients, so can the AI in this workspace.

That rarely matters for a single company with one site. It matters when one sign-in covers more than this workspace should, for example:

  • Agencies. Your Sitebulb, Semrush or Webflow login may hold every client’s projects. Connected to one client’s workspace, the AI could read the others too.
  • Several brands or sites. A company running separate workspaces per brand may share one account across them.

In those cases, connect with an account, or a workspace inside the provider, that holds only what this workspace should see. Some providers let you narrow access while you sign in, such as choosing which sites or bases to share. Where they do, the integration’s Source note says so, and it’s worth picking only what you need.

Google Analytics, Google Search Console, Google Ads and Bing Webmaster Tools work differently. After you connect, you choose the property, site or account to use, and Growth Method refuses any request for one you didn’t choose, even if your sign-in can reach it.

If a connection stops working

An integration’s access can stop working, usually because its sign-in has expired or someone revoked it at the provider’s end. When that happens, the AI can no longer use its tools.

You don’t have to watch for this. Growth Method checks your connected integrations every day, and also notices the moment one fails while the AI is using it. Either way, it marks the integration as disconnected and emails your team’s managers, so someone can reconnect it before it holds up your work.

A problem at the provider’s end, such as the service being down, overloaded or slow to answer, doesn’t count. It says nothing about your sign-in, so Growth Method leaves the integration connected and checks it again the next day.

Sometimes the sign-in still works but the integration’s calls fail, for example when the account behind it has lost access to the data. The integration stays connected, and its row in Settings, then Integrations, shows a muted note with the date its last call failed and the date it last worked. Hover over the note to see the error. The note goes once a call succeeds. If it stays, check the account with the provider, or reconnect the integration.

To reconnect:

  1. Go to Settings, then Integrations. The affected integration is marked as disconnected, with a short note on what went wrong.
  2. Select Reconnect next to it. What happens next depends on how it connects:
    • Sign in with the provider. You go through the provider’s sign-in again and, if the integration asks you to choose a property or site, you pick it again.
    • Enter an API key. The integration’s edit panel opens. Paste the new key and select Save changes. Growth Method checks the key works before it shows the integration as connected, and tells you if it doesn’t.

Any manager on the team can reconnect an integration, not only the person who first connected it. Everything that uses the integration, such as a goal’s metric or an agent’s tool choices, carries on as before once it’s back. The AI can use its tools again as soon as it reconnects.

Chat with one integration

A chat normally has every connected integration’s tools available. When you want the AI to focus on a single source, you can start a chat scoped to only that one.

  1. Go to Settings, then Integrations.
  2. Find the integration you want and select the chat icon next to it.
  3. A chat opens using only that integration’s tools, plus the always-on built-in Growth Method ones.

The chat keeps this scope even if you reload or reopen it later, and it lists what it’s using at the top. Because fewer tools are loaded, a focused chat uses less context and costs less to run. Anyone on your team can start one, not only managers.

Use your integrations from Claude or ChatGPT

An AI app you have connected to Growth Method, such as Claude or ChatGPT, can use the integrations on this page with your workspace’s connection. It is off until a manager switches it on.

  1. Go to Settings, then Integrations.
  2. Below the list, switch on Let AI apps use these integrations through MCP.

Tools you have switched off stay off, and anything that could change something follows the same approval rules as Growth Method’s own chat: it waits until the person who asked approves it. Switch it off again at any time and connected apps lose that access straight away. See MCP server for how it works and the limits that apply.

Tools

Tools are the actions an integration gives the AI, such as reading your analytics, searching your site, or creating a page. Each integration’s panel lists its tools, and you choose which ones the AI can use by switching them on or off. The field shows All tools when every tool is on, or how many are, such as 12 of 15 tools.

Some integrations expose hundreds of tools. Use Select all to turn every tool on, or the clear button on the field to turn them all off, then switch on only the few you need.

The AI fetches a tool only when it needs one, so a tool you leave on costs nothing until it’s used. Switch tools off to control what the AI is allowed to do, rather than to keep a chat cheap. The built-in Growth Method tools are always on and can’t be switched off.

Tool safety hints

Not every tool is equally safe to run without asking. A tool that only reads data is harmless to run automatically. A tool that changes or deletes something is not.

Growth Method uses each tool’s safety hints, plus an independent check on requests the hints don’t settle, to decide what runs on its own and what asks you first. See How approvals work.

Each approval appears as a card in the chat with Approve and Reject buttons. The AI pauses until you decide, and cards don’t expire: one still works when you come back hours or days later. A few things to expect:

  • Approving continues the chat. The action runs and the AI picks up where it left off, using the result.
  • Rejecting does too. The AI knows you declined, so it adjusts its approach or asks what you would prefer.
  • Several actions at once. If the AI proposes more than one action together, answer each card. The AI continues once every card has an answer.
  • Old cards retire themselves. If the chat moves on before you decide, the card changes to “Not actioned” and the action never runs. If you still want it, ask the AI again and a fresh card appears.

This matters most for a custom agent running on its own: unless you’ve set it to act autonomously, it only uses a tool unattended when the tool is marked read-only or the independent check clears the request, and asks you to approve anything else. The panel tells you how many of an integration’s tools only read data, and names the ones that may ask first: those that make changes, those that can delete or overwrite, and those the vendor hasn’t labelled. An autonomous agent runs those without asking, so it’s the list to check before you set one to act autonomously.

Editing files

Some integrations store files, such as a GitHub repository holding your website’s pages. When you ask the AI to change something in one of those files, it edits only the part you asked about instead of rewriting the file from scratch.

This matters for two reasons:

  • Long files work. Rewriting a whole page means the AI retyping every word of it. On a long article it can run out of room part-way through, and the change is lost.
  • You can see what you’re approving. The approval shows just the lines being changed, rather than the entire document.

You don’t need to switch anything on beyond the integration’s own file tools, which the AI uses to fetch the file and save it back.

A few things to expect:

  1. Be specific about what to change. The AI finds the text you’re changing by matching it exactly. If the same wording appears more than once in the file, it asks you which one you meant rather than guessing.
  2. Several changes go together. Asking for a title and a description change at once produces a single edit and a single approval.
  3. Creating a new file is different. There’s nothing to edit yet, so the AI writes the whole file. Very long new files can still run out of room, so it’s worth creating those yourself and asking the AI to edit them afterwards.

If someone else changes the same file between the AI reading it and you approving, the save is rejected rather than overwriting their work, and the AI tells you so it can try again.

Server instructions

Some integrations send the AI their own guidance on how to use their tools, written by whoever built the integration. Growth Method shows this to you read-only under Server instructions, so you can see exactly what the integration is telling the AI. If an integration provides none, the panel says so.

External websites

Some integrations mention external websites in the guidance they send the AI, often links to a provider’s own documentation. Growth Method scans that guidance and lists any it finds under External websites, so you can see what an integration points the AI at. If it mentions none, the panel says so.

Seeing a website here doesn’t mean the AI visits it. The AI can only open a web address if you’ve connected a separate tool that browses the web, which it doesn’t do on its own. The note is there for transparency, so nothing an integration references is hidden from you.

Source

Source tells you where an integration and its tools come from: built into Growth Method, a remote service at a web address, or a local server. Use it to check you recognise and trust the provider before you rely on its tools.

How this maps to the MCP standard

Many integrations connect through the Model Context Protocol (MCP), an open standard for connecting AI to external tools. The terms above line up with it directly:

  • Tools are MCP tools.
  • Server instructions are the MCP server’s instructions.
  • Tool safety hints are MCP tool annotations. The standard defines four:
    • readOnlyHint: the tool only reads, and changes nothing.
    • destructiveHint: if the tool does change something, the change may be permanent or irreversible (deleting, rather than adding).
    • idempotentHint: running the same call again has no extra effect.
    • openWorldHint: the tool reaches out to the wider world, like an external web API, rather than staying in a closed, local domain.

Growth Method acts on the first two, read-only and destructive, to decide what runs automatically and what needs your approval. idempotentHint and openWorldHint are part of the standard but don’t currently change how Growth Method behaves.

Some integrations, including the built-in Growth Method tools and certain analytics integrations, use built-in tools rather than MCP. These hints don’t apply to them, and their panel says so.

One important point from the standard: these are hints, not guarantees, and the standard tells apps to treat them as untrusted. That’s why Growth Method errs on the side of caution. A tool that isn’t clearly marked read-only is never assumed safe: each request is checked first, and anything the check isn’t confident about asks you. See How approvals work.

Next steps