> For the complete documentation index, see [llms.txt](https://docs.elevateab.com/elevate-helpcenter/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.elevateab.com/elevate-helpcenter/mcp/let-your-ai-assistant-make-changes.md).

# Let Your AI Assistant Make Changes

Give Claude, ChatGPT, or another MCP client permission to create, launch, pause, resume, and end experiments on your Shopify store with Elevate A/B Testing.

By default, every MCP connection is read-only. You can also give a connection permission to make changes, so your assistant can act on what it finds. For example, you can ask it to "create a Split URL test for the new sale page", or "pause the homepage test if its conversion rate drops".

A connection that can make changes can create, launch, pause, resume and end experiments. It can't edit an experiment's setup, delete one, or change anything in your Shopify admin. You choose which connections get it, and you can take it away by disconnecting.

***

### How do I let a connection make changes?

It depends on how the client signs in.

**Log in with Elevate (Claude, ChatGPT, Claude Code).** Ask your assistant to make a change. The first time, the client asks you to log in to Elevate A/B Testing again. The approval screen asks to read your data and make changes. Tick **Allow changes**, then approve. **Allow changes** is off unless you tick it, so approving without it keeps the connection read-only.

In ChatGPT, if the assistant says the connection is read-only, disconnect Elevate A/B Testing in ChatGPT, connect it again, and tick **Allow changes** on the approval screen.

**API key (Cursor, Gemini CLI, VS Code, other clients).** When you create the key, tick **Can make changes**. You can't change this later: to give an existing connection write access, create a new key with **Can make changes** ticked and disconnect the old one.

Connections that can make changes show **Can make changes** in the **Connections** list. See [Manage MCP Connections](https://docs.elevateab.com/elevate-helpcenter/mcp/manage-mcp-connections.md).

### What the assistant can change

These tools are added to the [read tools](https://docs.elevateab.com/elevate-helpcenter/mcp/mcp-overview.md#what-the-assistant-can-access):

| Tool                              | What it does                                                                                    |
| --------------------------------- | ----------------------------------------------------------------------------------------------- |
| `create_split_url_experiment`     | Creates a Split URL experiment, in Review                                                       |
| `create_product_image_experiment` | Creates a Product Image experiment, in Review                                                   |
| `create_custom_code_experiment`   | Creates a Custom Code experiment, in Review                                                     |
| `create_checkout_experiment`      | Creates a Checkout experiment, in Review. Shopify Plus, Plus trial, and development stores only |
| `launch_experiment`               | Launches an experiment that's in Review                                                         |
| `pause_experiment`                | Pauses a running experiment                                                                     |
| `resume_experiment`               | Resumes a paused experiment                                                                     |
| `end_experiment`                  | Ends a running or paused experiment. This is final                                              |

To build a Product Image or Checkout experiment, the assistant uses the read tools `search_products` and `get_product_media` to find your products and their images.

Launch and resume work on Page experiments too, but the assistant can't create them. Multiple Pages experiments can't be changed through MCP yet.

### What happens before the assistant launches anything

**New experiments start in Review.** Creating an experiment doesn't change your storefront. The assistant gets a preview link for each variation, so you can see it first.

**The assistant checks before it launches.** It runs a dry run first, which runs the same checks as the dashboard without changing anything. It's told to tell you the result and launch only once you agree.

**Custom code needs your confirmation.** Before a Custom Code experiment goes live, the assistant shows you the code and paths it will run. It can launch only after a dry run made in the last 10 minutes, and only if the code hasn't changed since.

**Every change has a reason.** The assistant records what you asked for, in your words, and it appears in the experiment's history.

**Checkout experiments need one step from you.** After launch, place Elevate A/B Testing's blocks in your checkout editor. The assistant reminds you, because no tool can do it.

### Where to see what your assistant changed

The MCP page lists **Changes made through MCP**: every change a connection made or tried to make, newest first, with the time, who made it, the action and its reason, the experiment, and the result. Dry runs and invalid requests aren't listed.

### Is it safe to let an AI assistant make changes?

**You decide per connection.** Read-only stays the default. Only connections you approved with **Allow changes**, or keys you created with **Can make changes**, can change anything.

**Launching changes what shoppers see.** Treat a connection that can make changes like a team member with access to your experiments. Ask it to show you a dry run and the preview links before it launches.

**Treat a write key like a password.** Anyone holding an MCP key created with **Can make changes** can launch experiments on your store. If a key is exposed, disconnect it straight away.

### FAQs

#### What happens if a read-only connection tries to make a change?

Nothing changes. The assistant is told the connection is read-only and how to get write access.

#### Can the assistant undo a change?

It can pause a running experiment, or resume a paused one. It can't undo an end: an ended experiment can't be resumed.

#### Do API changes and MCP changes share limits?

Yes. Both use your store's [rate limit](https://docs.elevateab.com/elevate-helpcenter/developer-reference/external-api/rate-limits.md), and each change costs 5 points.

#### Can I make the same changes from my own code?

Yes, with an External API write key. See [External API Write Access](https://docs.elevateab.com/elevate-helpcenter/developer-reference/external-api/write-access.md).
