pagent: the A/B testing MCP server for coding agents

@pagent/mcp connects Claude, ChatGPT, Claude Code, Codex, Cursor and any other MCP client to pagent, so your agent can run A/B tests on your live website. Set it up in one command:

npx @pagent/mcp init

This finds the MCP clients on your machine and points them at pagent's hosted MCP server. Nothing else to copy, no config format to choose.

What pagent is

pagent is an AI A/B testing agent, an AI Experiment Engineer for your website. It reads your pages, writes the hypothesis, builds the variation as real code on your design system, runs the test on your live site with Bayesian statistics, and learns what to try next. Your team reviews each variation before it goes live. pagent adds to your testing stack; it does not replace it.

With this MCP server your agent joins that loop: it reads results, captures the next hypothesis, and starts the next test from the chat you already use.

Related MCP server: Kumiki A/B

Can Claude or ChatGPT run A/B tests on my live website?

Yes, once they are connected to an experimentation platform through MCP. A language model on its own cannot serve variations to your visitors or measure conversions. With the pagent MCP server, Claude, ChatGPT or your coding agent can start, pause and stop tests on your live site and read their results, while pagent builds the variations, splits the traffic and runs the statistics.

Things you can ask once connected:

• "Which of our running A/B tests is winning, and how sure are we?"

• "Compare last month's pricing page tests and tell me what changed on the page."

• "Capture a hypothesis: a shorter signup form will raise trial starts. Turn it into a test."

• "Pause the checkout test and roll it back."

What your agent can do

With read access:

• List running and completed tests and read their Bayesian results

• Compare tests and inspect what each variation changes on the page

• Read pages, audiences, conversion goals, personas and guidelines

With write access:

• Capture a hypothesis and turn it into a test

• Start, pause, resume, queue, stop, evolve or roll back a test

• Create pages and conversion goals, manage audiences

You choose read or write, and which organisations the client may see, during sign-in. You can revoke a connection at any time in your pagent settings.

Usage

npx @pagent/mcp init                 # detect your clients and connect them
npx @pagent/mcp init --client cursor # connect one client
npx @pagent/mcp init --local         # write a project config, not a user one
npx @pagent/mcp init --dry-run       # show what would change, write nothing
npx @pagent/mcp print                # print the config without writing
npx @pagent/mcp list                 # list supported clients

Supported clients

Claude and ChatGPT are configured in their own interface, so this CLI prints the steps rather than editing anything. Run npx @pagent/mcp print --client chatgpt.

Existing servers in your config are left alone. If a config file cannot be parsed, for example because it has comments, the CLI prints the entry for you to paste instead of rewriting the file.

Connecting by hand

Any MCP client can use the hosted server directly:

• Endpoint: https://app.pagent.ai/api/mcp

• Transport: Streamable HTTP

• Authentication: OAuth 2.0, or a personal token from AI client access

Before you connect

You need a pagent account with owner or administrator access to the organisation, MCP enabled for it, and a website with the pagent script installed so there are tests and visitor data to read.

Links

• pagent MCP server for A/B testing

• Setup guide

• The AI A/B testing agent

License

MIT. pagent itself is a commercial product; this CLI only writes local configuration files.