Two directions
Agent-native in both directions
pagent is the agent. It defines hypotheses, builds and checks real-code variants on your design system, and runs statistically rigorous tests on your live site. And pagent is a tool your agent can drive. Claude, ChatGPT, Codex, Cursor and any MCP client connect over OAuth to read results, capture hypotheses, create and start tests, and manage audiences and goals.
Your agent gives the goal. pagent does the work. Your team reviews each variant before it goes live.
Capabilities
What your agent can do
With read access
- List running and completed tests and read their Bayesian results in plain language
- Compare tests and inspect what each variant changes on the page
- Read pages, audiences, conversion goals, personas, strategies and guidelines
- List the websites and organisations the connection may see
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 and their pages
- Update personas, strategies, keywords and guidelines that steer the agent
You choose the level per connection during OAuth, and which organisations the client may see. You can revoke a connection at any time from your settings.
Setup
Connect in one step
Claude
Open pagent in Claude's connector directory, select Connect, sign in, and choose organisations and permissions.
ChatGPT
Turn on Developer mode under Settings, Security and login. Then open Plugins, select the plus button, name the plugin pagent, enter the endpoint as the server URL, keep OAuth as the authentication, and select Create. Sign in with pagent, choose organisations and permissions, and allow access. To use it, open Plugins, pagent, Try in chat, or type /pagent in the composer.
Claude Code
claude mcp add --transport http pagent https://app.pagent.ai/api/mcpCodex
codex mcp add pagent --url https://app.pagent.ai/api/mcp
codex mcp login pagentCursor
{
"mcpServers": {
"pagent": { "url": "https://app.pagent.ai/api/mcp" }
}
}VS Code, Windsurf, Zed and personal tokens: the setup guide.
Prompts
What to ask once connected
- Which tests are running on shop.example.com, and which one is ready to ship?
- Turn this hypothesis into a test on /checkout: fewer form fields above the fold raise completed orders.
- Compare the two hero tests from August and explain the result to my team.
Name the website in each request. A connection does not remember a selected site between messages.
The difference
What pagent does that your agent does not
An MCP server can create a test from selectors your agent authors. pagent goes further. pagent reads the page, writes the hypothesis, builds the variant as real code with your components and your tone, checks it on desktop and mobile, and runs it with Bayesian statistics, sequential correction and a sample-ratio check. Winners become the new baseline and shape the next hypothesis.
How the AI A/B testing agent works, and the wider field in the best AI A/B testing tools in 2026, by use case.
FAQ
Frequently asked questions
Is there an A/B testing MCP server I can use with Claude or ChatGPT?
Yes. pagent runs a hosted MCP server at https://app.pagent.ai/api/mcp over streamable HTTP with OAuth 2.0. It is listed in Claude's connector directory and in the official MCP registry, and it works in Claude, Claude Code, ChatGPT, Codex, Cursor, VS Code, Windsurf, Zed and any other MCP client.
Can Claude run A/B tests on my live website?
Yes, with the pagent connector. Add pagent from Claude's connector directory, sign in, and ask Claude to turn a hypothesis into a test. pagent builds the variant as real code on your site, checks it on desktop and mobile, and runs the test with Bayesian statistics. Claude reads the results and captures the next hypothesis.
Does my agent have to write the variant code?
No. Your agent gives the goal and the hypothesis. pagent writes the variant as real code on your design system, checks it on desktop and mobile, and ships it through the script tag already on your site. That is the difference to an MCP server that only creates a test from selectors your agent has to author.
What is the difference between read and write access?
Read access lets an agent list tests, read Bayesian results, compare variants and inspect pages, audiences, goals and personas. Write access adds the test lifecycle: capture hypotheses, create and start tests, pause, stop, evolve or roll back, and manage audiences and goals. You choose the level per connection during OAuth.
Is pagent agent-native?
In both directions. pagent is the agent: it defines hypotheses, builds and checks real-code variants, and runs statistically rigorous tests on your live site. And pagent is a tool your agent drives over MCP. Most tools offer one of the two.
Does anything ship without a review?
Not without your team. Every variant goes through review in the dashboard, by email, or from the chat you already use. pagent does not add a second confirmation screen for MCP writes, so keep write confirmation on in your client and ask it to show planned changes before it acts.
What do I need before I connect?
A pagent account with owner or administrator access to the organisation, MCP enabled for that organisation, and an onboarded website with the pagent script installed so there are tests and visitor data to read.