Get started
MCP server

Work in Cuttl from Claude, ChatGPT or any AI assistant

Cuttl’s MCP server lets AI assistants read your heatmaps, recordings and test results, and do the work you’d do in the dashboard, from building A/B tests to triaging issues. Anything that changes your site waits for a person to approve it.

Get startedHow to connect
Sign in with OAuthUse your Cuttl account. No API key to paste.
Works where you workClaude, ChatGPT, Cursor, VS Code and other MCP clients.
Approval before changesNothing goes live until a person says yes.
Every call loggedTool, inputs, result and who approved it.

Connect in a minute

Sign in with your Cuttl account. The Connect AI page in Cuttl has a button for each assistant.

Claude

Desktop app or web. The button copies the server address and opens the add-connector window.

ChatGPT

On paid plans, in developer mode: add a custom MCP server and sign in with OAuth.

Cursor

One-click install link.

VS Code

One-click install link.

Claude Code

One command in your terminal.

claude mcp add --transport http cuttl https://app.cuttl.io/mcp

Gemini CLI and others

Any tool that supports MCP sign-in.

Server addresshttps://app.cuttl.io/mcp

API keys are still available for scripts and automations.

Ask anything about your visitors

Assistants read the same analytics you see in Cuttl, with recordings as summaries and replay links, and every analysis tool takes the same segment filters: device, country, channel, UTM, new or returning visitors and more.

  • Overview, session and visitor search with summaries and replay links, and one visitor across all their visits
  • What a page’s heatmap shows: clicks and scroll depth
  • Rage, dead and error clicks, quick backs, and detected issues with lost conversions
  • Funnels, forms, feedback, goals, site health, traffic quality, Change Impact and past test learnings

Build A/B tests the way a person would

The assistant reads the page first. A tool lists its headings, buttons, links, fields, images and text, with the exact selectors the visual editor uses, taken from a real recorded visit. Then it sets up the test.

  • A/B, split URL, multi-page and multivariate tests, and rollouts
  • Page rules, variants, goals, audience, trigger, traffic split and a pre-registered minimum detectable effect
  • Text, HTML, style, attribute, hide, move, insert, CSS and custom code changes
  • Every change is checked against a recorded visit of the page, and misses are flagged so the assistant fixes them first
  • Preview links, then it asks before launching, pausing, ending, picking a winner or rolling out
Learn more about A/B testing

See the variant without leaving the chat

In assistants that support MCP Apps, such as Claude, the test preview opens inside the conversation: original and variant side by side, with a divider you can drag. It opens on the first change, outlines what changed and lets you switch between variants. Elsewhere, the assistant gives you preview links.

Claude · connected to Cuttl

Draft a test that shows delivery costs on the product page, before the cart.

Variant B is saved as a draft. All 3 changes matched the page. Here’s the preview:

Variant BVariant C3 changes · opened on change 1
Fernwood Home

Stonewashed linen duvet cover

$129

Delivery calculated at checkout

Add to cart
Fernwood Home

Stonewashed linen duvet cover

$129In stock · ships tomorrow

Free delivery over $75 · arrives in 2 to 4 days

Add to cart · free delivery
OriginalVariant B

Shall I launch it to half your visitors? A launch needs approval in Cuttl first.

Set up and run Journey Insights

Setup work goes through the chat too. Goals are counted over your past data straight away, so a new goal has numbers the moment it exists.

  • Goals, funnels and saved segments
  • Notes on heatmaps and on moments in recordings, and starred recordings
  • Issue triage: status, owner and “not fixed”
  • Deploy markers, page groups and alert thresholds
  • Feedback widgets, surveys, and uptime, API and browser journey monitors

You stay in control

Assistants work through approvals, permissions and a log, so larger teams can let them help without handing over the keys.

1

The assistant asks

It drafts the change, such as a new test or a launch, and sends it to Cuttl for review.

2

A person approves

In Cuttl, under AI approvals, with the same roles as the dashboard.

3

It goes live, on record

The change runs, and the call is saved in the AI action log with who approved it.

Permissions per connection

Read is always on. Write-draft, write-setup, write-launch and auto-approve-low-risk are switched on per connection by an admin.

The roles you already use

Analysts approve content, admins approve site settings and high-risk actions, and launches need launch permission.

A full action log

Every assistant call is logged: the tool, its inputs, the result and who approved it.

Always needs a person

  • Launching a test
  • Edits to a live test
  • Custom code
  • Deleting anything
  • Anything visitors will see

Low-risk drafts can auto-approve if you turn that on.

Never available to assistants

  • Privacy, consent and retention settings
  • Integrations and webhooks
  • Share links
  • Deleting visitor data
  • Managing the team or API keys

Personal data in page text is masked, and a connection only sees its own workspace.

Journey Insights

Heatmaps, session recordings, funnels and the issues Cuttl finds for you.

Explore

Web Experiments

A/B, split URL and multivariate tests, a visual editor and Change Impact.

Explore

MCP questions

MCP (Model Context Protocol) is an open standard that lets AI assistants connect to other software. Cuttl runs an MCP server at https://app.cuttl.io/mcp, so assistants such as Claude and ChatGPT can read your Cuttl data and, with approval, do work in Cuttl.

Claude (desktop app and web), ChatGPT (paid plans, developer mode), Cursor, VS Code, Claude Code and Gemini CLI, plus any other tool that supports MCP sign-in.

No. You sign in with your Cuttl account through OAuth, so there is no key to copy or paste. API keys are still available for scripts.

No. Nothing changes until someone approves it in Cuttl, unless you choose to auto-approve low-risk drafts. Launches, edits to live tests, custom code, deletes and anything visitors see always need a person.

The analytics of the workspace it is connected to, such as sessions, heatmaps, issues, funnels and test results. Personal data in page text is masked, and a connection only sees its own workspace.

It can’t change privacy, consent or retention settings, integrations or webhooks, create share links, delete visitor data, or manage the team or API keys.

Yes. Cuttl has one plan, and the MCP server is part of it along with every Journey Insights and Web Experiments feature.