Why this guide is different

Every other setup guide in this series is built around one tool that most readers can use for free: GA4, Search Console, Clarity, Tag Manager. A/B testing doesn’t work like that. There’s no single dominant free platform, the tool you use is often decided by your business or client, and MCP support varies widely between vendors.

So this post does three things. It compares the testing platforms that currently offer MCP servers. It walks through connecting two of them to Claude, Convert and Unbounce, as worked examples. And it shows what becomes possible once your testing platform, Tag Manager and GA4 are all connected to the same AI assistant. It’s deliberately light on experimentation theory: our Introduction to CRO series, starting on 8 November, covers designing and running experiments in depth.

Which A/B testing platforms have MCP servers?

This is a fast-moving area, so treat the table below as a snapshot from October 2026 and check each vendor’s own documentation before choosing. Free tiers are listed separately because they often have different limits from the paid plans.

Platform	MCP server	Connection	What it can do via MCP	Free option
Convert	Official	Hosted (OAuth) or local (API key)	Read results, audit and QA experiments, create draft A/B and Split URL tests, manage goals and audiences	Free trial only
Unbounce	Official, included on all plans	Claude plugin (OAuth)	Build landing pages, create variants, split traffic, Smart Traffic, publish, page stats	Free trial (A/B testing needs the Experiment plan or higher)
Wingify (VWO)	Official	Hosted (OAuth)	Read campaign reports, heatmaps and recordings, sample size calculation, create A/B and Split URL tests, start and pause tests	Free trial
VWO Starter (free plan)	Via the Wingify server above	Hosted (OAuth)	As above, within the plan’s features	VWO has offered a free Starter plan, but check current availability since the Wingify rebrand
Wingify Feature Experimentation	Official, open source	Local (API key)	Feature flags and server-side experiments, aimed at developers	Depends on Feature Experimentation plan
GrowthBook	Official	Aimed at AI coding tools	Feature flags and experiments in code	Free cloud tier (up to 3 users) and free self-hosting
PostHog	Official	Aimed at AI coding tools	Product analytics, feature flags and experiments	Generous free tier; experiments need PostHog Cloud
Kameleoon	Reported	Hosted (OAuth), reported	Reported to cover experiments, goals, segments and flags. We couldn’t confirm this from Kameleoon’s own documentation.	No free plan found
Optimizely	Commerce only	—	Optimizely’s MCP server currently covers commerce, not experimentation	No free plan
AB Tasty	None found	—	Now part of Wingify, so this may change	No free plan

The pattern is clear: visual and marketer-focused tools (Convert, Unbounce, VWO) offer hosted MCP servers you sign in to, while developer-focused platforms (GrowthBook, PostHog, Feature Experimentation) aim theirs at AI coding tools such as Claude Code.

Worked example 1: connecting Convert to Claude

Convert offers a hosted connector, which is the simplest route. You don’t need API keys or a config file.

1. In Claude (desktop app or claude.ai), open Settings → Connectors.
2. Click Add custom connector.
3. Give it a name, for example Convert, and paste the connector URL: https://mcp.ai.app.convert.com/mcp
4. Click Add, then Connect.
5. A Convert sign-in page opens. Sign in, then choose which Convert accounts and projects Claude can access. Only select the ones you need.
6. Start a new chat and test it: “How many experiences do I have running in Convert right now?”

Prefer a local setup? Convert also publishes a local server that uses your Convert API key and secret. It has three access levels, set with the TOOLS_FOR_CLIENT variable: reporting (the default), readOnly, and all, which adds create, edit and delete actions. Convert’s MCP help article has the full configuration. Start with reporting and only move to all when you need it.

Worked example 2: connecting Unbounce to Claude

Unbounce installs as a Claude plugin, which bundles the MCP connector with Unbounce’s own skills. These are the steps from Unbounce’s documentation for the Claude desktop app:

1. Open Customize → Plugins, click Add and choose Add marketplace.
2. Choose Add from a repository, type unbounce/agent-plugins, select it from the dropdown and click Sync. Leave Sync automatically switched on.
3. Under Plugins → Personal → agent-plugins, click the Unbounce mcp tile, then Install.
4. Click Manage, open the Connectors tab and click Install on the unbounce connector. In the dialog that opens, click Add without changing anything.
5. Click Connect, sign in to Unbounce in your browser and authorise access.
6. Test it: “What version of Unbounce MCP am I using and what tools do I have access to?”

Two things to know before you start. First, pages built through the MCP are code-based, so they can’t then be edited in Unbounce’s visual builder; all further changes go through your AI assistant. Second, pages built in the classic builder can’t be edited through the MCP directly, but they can be copied into an MCP-managed page. If your organisation manages your Claude account, an admin may need to approve the plugin first.

What Claude can do once your testing platform is connected

Get more out of your experiments

With the testing platform connected alongside GA4 and Clarity from earlier parts, Claude can go further than the platform’s own reports:

• “Summarise every experiment that finished in the last quarter: hypothesis, result, uplift and confidence. Group them by page type.”
• “For our pricing page test, compare the variant’s results by device, then check Clarity session recordings for the same period to explain the difference.”
• “Run an active health check on my running experiments and flag any with uneven traffic splits, missing goals or results that look too good to be true.”

Build an experiment and its tracking at the same time

This is where connecting several servers pays off. In Part 6 we connected Claude to Google Tag Manager. With both connected, a single request can cover the whole setup:

“Create a draft A/B test in Convert for the homepage hero: control plus one variant with the headline ‘Get a quote in 60 seconds’. Add a goal for clicks on the quote button. Then, in a new GTM workspace called ‘Hero test tracking’, set up a trigger and GA4 event tag that send the experiment ID and variant name to GA4. Don’t start the test or publish GTM. Show me everything you’ve created.”

Claude builds both halves, then gives you a single summary to review. You then check the experiment in the testing platform’s preview and the tracking in GTM’s Preview mode before anything goes live.

Launch landing page tests faster

With Unbounce, Claude can build the test itself: “Create a second variant of the webinar page with a shorter form and the testimonial moved above the fold. Split traffic 50/50 but leave the page unpublished.” Changes to content, variants and traffic stay as drafts until you ask Claude to publish.

Safeguards: keeping Claude’s experiments safe

An experiment that launches broken, or with tracking missing, wastes traffic you can’t get back. These habits help:

• Start with read-only access. Use Convert’s reporting level, Wingify’s Browse permission, or authorise only the projects you need. Widen access only when you’re ready to build.
• Always create drafts. Tell Claude explicitly not to start tests, publish pages or publish GTM, and make launch a separate instruction.
• Ask for a pre-launch QA plan. For example: “Draft a QA checklist for this test, covering targeting, goals, every device, and confirming the GA4 events fire for each variant.”
• Check the numbers before you launch. Ask Claude to estimate the sample size and duration from your GA4 traffic and current conversion rate, and question tests that can’t reach a result in a reasonable time.
• Watch the first days. Ask for a check after 24 hours: is the traffic split even, are both variants recording conversions, and are there any errors?

Going deeper

This post covers the connection and the workflow. Deciding what to test, prioritising ideas, designing experiments properly and interpreting the results are all covered in our Introduction to CRO series (coming soon). It includes dedicated parts on ideation and experimentation, and a part on using AI for CRO.

Next up: Part 9 — WordPress MCP: Connect Claude to WordPress, so Claude can make the fixes it finds directly on your site.