Guide · MCP

A/B testing MCP servers

An MCP server lets your own AI agent operate an A/B testing tool. Of the tools we track, 21 document one (fully or partly), plus Duelpilot, whose MCP server is how the whole product is run.

Last reviewed 2026-10-17Markdown version

Which A/B testing tools have an MCP server

Tool	MCP server	Your agent designs variants	Agent can launch
Duelpilot	Yes Remote MCP server with 48 tools; OAuth for claude.ai and ChatGPT, bearer token elsewhere. source	Yes Your agent designs 2–4 variants per idea, each with a preview and a device check. source	No A person approves every launch; an owner can switch on an autopilot policy with guardrails, never an agent. source
Amplitude Experiment	Yes mcp.amplitude.com/mcp. source	Yes create_experiment with custom variants, create_flags, update_experiment. source	Unknown No dedicated start tool documented; update_flag covers deployments. source
Convert	Yes Hosted OAuth connector and the @convertcom/mcp-server package: 16 top-level tools with 113 callable actions when full access is enabled. source 1, source 2	Yes With full access, the agent can create, clone and update experiences and variations; new work is created as a draft. source	Partly Status changes are possible with full access, but activating or pausing requires explicit confirmation; the local server starts read-only. source
Crazy Egg	Yes mcp.crazyegg.com/mcp with OAuth. source	Unknown The documented tools are chat and harness tools; A/B-test actions are not described. source	Unknown No test-launch tool documented. source
Eppo (Datadog Experiments)	Yes Datadog MCP Server with the experiments toolset enabled (off by default). source	Yes create_experiment, create_datadog_feature_flag, link_feature_flag_to_experiment. source	Yes start_experiment, conclude_experiment and cancel_experiment tools; no approval gate is documented. source
Flint	Yes Remote MCP server at mcp.tryflint.com/mcp, added as a Claude connector. source 1, source 2	Partly Claude creates and edits pages, not test variants. source	Partly publish_site deploys a page to production; there are no experiment tools. source
GrowthBook	Yes Hosted at mcp.growthbook.io/mcp or local via @growthbook/mcp. source	Yes growthbook_api_write can make any write call the REST API allows, including creating experiments. source	Yes Optional: the API write tool can start experiments; confirmation can be required per API key, and the bundled skills describe draft, review, publish. source
Humblytics	Yes Hosted MCP at mcp.humblytics.com (Streamable HTTP), included on every plan. source 1, source 2	Partly create_split_test drafts a test, but every write is a proposal the user must explicitly approve. source	No No start tool; the page says nothing goes live until the user says so and the agent must never promote a winner itself. source
Intelligems	Yes Included in every paid plan; works with Claude, ChatGPT, Gemini and Grok. source	Yes create_experience and update_experience cover variations, targeting and traffic. source	Yes experience_action can start, pause or end a test; new experiences are drafts and the AI client asks for confirmation before running them. source
Kameleoon	Yes mcp.kameleoon.com/mcp with OAuth; Claude, ChatGPT, Claude Code, Codex and Cursor. source 1, source 2	Partly experiment_create makes a draft with common JS/CSS; no documented tool to create Web variation code. source	Yes experiment_lifecycle_update can start, pause, resume, stop or delete; only stop and delete require confirmation. source
LaunchDarkly	Yes Hosted MCP at mcp.launchdarkly.com with OAuth. source	Yes create-flag, create-experiment, update-experiment. source	Yes Tools such as toggle-flag, update-rollout and start-experiment-iteration exist; where an environment requires approval, the agent can only file an approval request for a reviewer. source
Mida	Yes mcp.mida.so/mcp with OAuth and dynamic client registration; launched May 2026. source	Yes Describe the test and Mida builds the variant, targeting, goal and preview link. source	Yes Tests are drafts until launched, but launching and serving the winner are MCP actions with no separate dashboard approval. source
Optimizely	Yes Hosted remote MCP with Opti ID OAuth; requires Optimizely Agent Platform to be enabled. source 1, source 2	Partly Can create and update flags, experiments and audiences, but Web variation code (custom HTML, CSS, JS) is not supported through MCP. source	Unknown No documented start or publish action; the agent confirms in the client before committing changes. source
pagent	Yes app.pagent.ai/api/mcp, Streamable HTTP with OAuth or a personal token; works with Claude, ChatGPT, Codex, Cursor and others. source	Yes The agent can capture a hypothesis, turn it into a test and work with pages, audiences and conversion goals. source	Yes With write scope granted at OAuth the agent can start, pause, stop, roll back and mark winners shipped; pagent adds no second confirmation and relies on the client's confirmation setting. source
PostHog Experiments	Yes Official PostHog MCP (US and EU). source	Yes experiment-create, experiment-create-from-prompt, experiment-update. source	Yes experiment-launch, -pause, -resume, -end and -ship-variant; approval is left to the MCP client, and write tools can be removed with read-only mode. source 1, source 2
Statsig	Yes Official MCP server with experiment, gate and API tools. source	Yes experiment_create and gate_create. source	Unknown No dedicated start or ship tool is documented; update and destructive tools need a confirmation parameter and review policies still apply. source
Tailor	Yes Tailor MCP server for Claude, Claude Code and internal agents (Pro and above). source	Yes Agents build experiments and variants: headlines, CTAs, images, full sections, layout and scripts. source	Yes Optional: "Launch can be open to agents or reserved for humans"; agents can start, pause and ramp experiments when allowed. source
Unbounce (Smart Traffic)	Yes Included with every Unbounce plan. source	Yes "Create variants, allocate traffic, start or stop tests, and deploy winning variants." source	Yes The agent can start and stop tests and deploy winners; A/B testing needs the Experiment plan or higher. source
VWO (Wingify)	Yes Web-testing MCP at mcp.wingify.ai (14 tools) and a separate FME MCP for feature flags. source 1, source 2	Yes Create A/B and split-URL campaigns, clone and update configuration (not MVT). source	Yes With Publish/Admin permission an agent can start, pause or stop a campaign; no server-side approval step is documented. source
ABConvert	Partly In beta, opt-in via support, Growth plan and above. source	Yes Creates price, shipping, theme, template, URL, offer and checkout tests; visual-editor tests stay wizard-only. source	Yes Tools cover launch, pause, resume and end; drafts are created after the user confirms in chat. source
Optibase	Partly MCP on all plans, but read-only. source	No The MCP cannot create, edit or delete tests. source	No Read-only MCP. source
Webflow Optimize	Partly Webflow's MCP can read Optimize experiments and results (announced 15 Oct 2026); no write access. source	No Only read access is announced. source	No People create and launch in the Webflow UI. source
AB Tasty	Unknown No AB Tasty-specific MCP found; whether Wingify's MCP serves AB Tasty-platform accounts is not documented. source	Unknown Not documented for the AB Tasty platform. source	Unknown Not documented for the AB Tasty platform. source
Fibr AI	Unknown No MCP server or public API mentioned. source	Unknown Fibr's own AI generates variants; no route for a customer's agent documented. source	Unknown Not documented for external agents; in the platform, variants are approved with one click and winners can be auto-promoted. source
Coframe	No No customer MCP server for test management found; the site only exposes a WebMCP recommendation tool for visitors' agents. source	No Coframe's own AI creates variants; your own agent cannot. source	No Variations need the customer's approval before going live. source

As documented by each vendor. Each cell links to its source.

Connect Duelpilot's MCP server

The server is at https://duelpilot.com/api/mcp (Streamable HTTP, stateless). Pick what your client supports:

claude.ai, Claude Desktop and ChatGPT: add a custom connector with URL https://duelpilot.com/api/mcp and sign in (OAuth). You pick the organisation, sites and permissions; there is no token to paste.
Claude Code: run the command below with your Duelpilot token (the tools appear in a new session).
Cursor and other MCP clients: a remote server with URL https://duelpilot.com/api/mcp and the header Authorization: Bearer <token>.
No account yet? Your agent can open a workspace itself and hand you a claim link: instructions for agents.

claude mcp add --transport http duelpilot https://duelpilot.com/api/mcp --header "Authorization: Bearer $DUELPILOT_TOKEN"

What your agent can and can't do

48 tools cover installing and verifying the script, reading GA4 and page structure, proposing ideas, designing variants with previews and device checks, reading results and posting updates. Its only protective action is pausing a test. There is deliberately no tool that approves, launches, resumes, picks a winner, ships, changes policy or manages tokens: those stay with people in Duelpilot's board.

Try Duelpilot with your own agent

Give your agent one line: set up Duelpilot for my site using /agents.md. It opens a workspace, installs and verifies the script, and drafts the first test; you approve what goes live. Try your first test free or see pricing.

Questions

What is an A/B testing MCP server?

A Model Context Protocol server that exposes an A/B testing tool's actions (read results, propose a test, check a page) so an AI agent like Claude, ChatGPT or Cursor can use the tool directly.

Which A/B testing tools have an MCP server?

Duelpilot, ABConvert, Amplitude Experiment, Convert, Crazy Egg, Eppo (Datadog Experiments), Flint, GrowthBook, Humblytics, Intelligems, Kameleoon, LaunchDarkly, Mida, Optibase, Optimizely, PostHog Experiments, Statsig, Tailor, Unbounce (Smart Traffic), VWO (Wingify), Webflow Optimize, pagent, as documented by each vendor (some only partly; see the table).

Can an agent launch tests through Duelpilot's MCP server?

No. Agents propose and check; a person approves every launch in Duelpilot's board (or an owner's autopilot policy within guardrails).

Sources

Last reviewed 2026-10-17. Spot an error? Tell us.

Anthropic: Introducing the Model Context Protocol (25 Nov 2024) (checked 2026-10-17)
Agent-Native CRO for Claude & Codex | Humblytics (checked 2026-10-17)
Humblytics MCP (checked 2026-10-17)
Pagent MCP server for Claude, ChatGPT, Codex and Cursor (checked 2026-10-17)
Tailor MCP: Connect Your AI Agents to Website Experiments (checked 2026-10-17)
Coframe — AI landing pages that maximize revenue (checked 2026-10-17)
Optimizely Experimentation MCP server overview (checked 2026-10-17)
Optimizely Experimentation MCP server quickstart (checked 2026-10-17)
Optimizely Experimentation MCP server example prompts (checked 2026-10-17)
Kameleoon MCP server (checked 2026-10-17)
Connect AI assistants | Kameleoon (checked 2026-10-17)
Mida MCP server | Connect Mida to your AI agent (checked 2026-10-17)
Changelog | Mida (checked 2026-10-17)
Fibr AI (checked 2026-10-17)
Fibr AI Experimentation (checked 2026-10-17)
Webflow updates: MCP for Optimize, Analyze and AEO (checked 2026-10-17)
Convert MCP Server (checked 2026-10-17)
Convert MCP Server | Convert Support (checked 2026-10-17)
Connect Wingify with AI Tools using Wingify's MCP Server (checked 2026-10-17)
MCP Server | Wingify FME (checked 2026-10-17)
AB Tasty Docs (checked 2026-10-17)
PostHog MCP tools reference (checked 2026-10-17)
PostHog MCP FAQ (checked 2026-10-17)
Pricing & Plans | Intelligems (checked 2026-10-17)
Available Tools | Intelligems Docs (checked 2026-10-17)
Drive ABConvert from Claude with MCP (checked 2026-10-17)
What Claude can do for you | ABConvert Docs (checked 2026-10-17)
Optibase MCP: Ask AI About Your A/B Tests in Plain English (checked 2026-10-17)
Crazy Egg Model Context Protocol (MCP) Connector (checked 2026-10-17)
Unbounce MCP: Build Landing Pages in Claude & ChatGPT (checked 2026-10-17)
MCP Server | GrowthBook Docs (checked 2026-10-17)
MCP | Statsig Docs (checked 2026-10-17)
Datadog MCP Server Tools (checked 2026-10-17)
Amplitude MCP (checked 2026-10-17)
LaunchDarkly hosted MCP server (checked 2026-10-17)
Claude MCP | Flint Docs (checked 2026-10-17)
Flint MCP | Connect Claude to Your Landing Page Workflow (checked 2026-10-17)
Duelpilot instructions for AI agents (checked 2026-10-17)
Duelpilot overview (checked 2026-10-17)
Duelpilot pricing (checked 2026-10-17)