MCP server · Claude, ChatGPT, Cursor and more

Load testing, by asking your AI.

Your site is fine with a few visitors. Ask your AI what happens with 500. It writes the test, checks it in a real browser and tells you which page gave out first.

Connect your AI

One sentence in.
A tested site out.

An example conversation, shortened. Every assistant uses the same engine as the web app.

Your AI assistantUserLimit connected

Load test shop.example.com with 50 users: open the homepage, search for running shoes, filter by size 42 and add the first pair to the cart.

get_account_overview create_scenario validate_scenario prepare_test_run

The domain is already verified. I checked the click path with one free test session, and it passed. Each of the 50 users will:

1Open the homepage
2Search for “running shoes”
3Filter the results by size 42
4Open the first product
5Add it to the cart

Should I start the run? It sends real traffic to shop.example.com for about 5 minutes and uses 0.6 of your 3 free hours. It stops itself if more than 5% of visits fail.

Yes, start the test

start_test_run get_test_run

The test passed. All 50 users reached the cart. Search was the slowest step at 1.9 s on average, so that is the place to look first.

Score92
Avg page load640 ms
Server response210 ms
Errors0

The assistant offers to rerun with 300 users.

Works with

  • Claude
  • ChatGPT
  • Cursor
  • VS Code
  • Windsurf
  • Any MCP client
No scripts, no dashboards

Ask your AI assistant to load test your site

Your domains, scenarios and past runs live in UserLimit, so your assistant picks up where you left off.

First test

“Can my site handle 200 people signing up at once after our launch email?”

It adds the domain, writes the sign-up visit, checks it for free and sizes the test to your hours.

After a fix

“Rerun last week’s checkout test and tell me if it got faster.”

Every result is compared with the previous run: load times, errors and the slowest pages.

Make it a habit

“Run the catalog test every weekday at 7 and post problems to Slack.”

Schedules and alerts are set from the chat, so a slow deploy shows up the next morning.

Compared

A hosted MCP server for load testing, not another tool to run

MCP servers for k6 or JMeter let your assistant drive a load tool you install and host. UserLimit is the whole service behind one URL.

UserLimit MCP serverk6 or JMeter MCP server
What runsReal browsers: pages, scripts, clicks and formsProtocol requests to URLs and APIs
What you measureWhat visitors see: page loads, failed steps, slow pagesResponse times and status codes
SetupPaste one URL, sign inInstall the tool and the server, run load generators
Where the load comes fromUserLimit’s cloud, in the regions you pickYour machine or infrastructure you provide
SafetyVerified domains, signed confirmation and stop rules, enforced by the serverUp to the model and your setup
ResultsScore, slowest pages, comparison with the last run, shareable reportRaw output for the assistant to read

Need API or requests-per-second testing? A protocol tool is the better fit. UserLimit tests what people do in the browser.

For developers

19 MCP tools, from domain to report

Each tool is marked read-only or destructive, so your client knows when to ask for approval.

Show all 19 tools

Read-only or reversibleDestructive: your client asks first

Domains

Add a site and prove it’s yours.

  • get_account_overview
  • add_domain
  • verify_domain
  • remove_domain (destructive, asks first)

Scenarios

Write a visit as Playwright and check it in a real browser, free.

  • create_scenario
  • validate_scenario
  • get_scenario
  • update_scenario (destructive, asks first)
  • delete_scenario (destructive, asks first)

Tests

Sessions, regions and schedules, with usage estimates.

  • preview_test_settings
  • get_test
  • get_account_usage
  • create_test (destructive, asks first)
  • update_test (destructive, asks first)
  • delete_test (destructive, asks first)

Runs

Confirm, run, follow along and share the report.

  • prepare_test_run
  • get_test_run
  • start_test_run (destructive, asks first)
  • stop_test_run (destructive, asks first)

Parameters and the full flow are in the MCP setup guide.

Get started

Connect the MCP server to Claude, ChatGPT, Cursor or VS Code

Pick your assistant. No API key: your client opens the UserLimit sign-in, where you can also create a free account with 3 free hours. Then just ask, and nothing runs until you say yes.

Step-by-step setup and troubleshooting are in the MCP setup guide. Tests use virtual-user hours, priced per hour after the free ones. Prefer the web app? Everything your assistant does is also in the dashboard.

FAQs

What is the UserLimit MCP server?

It is a remote Model Context Protocol server that lets AI assistants such as Claude, ChatGPT, Cursor or VS Code use your UserLimit account. Your assistant can verify a domain, write and check a scenario, size the test, run it after you confirm, and read back the results. It uses Streamable HTTP with OAuth sign-in, so there is no API key to copy.

How do I connect my AI assistant?

Add https://userlimit.com/mcp as a remote MCP server in your client. In Claude that is Settings, Connectors, Add custom connector; in ChatGPT, Cursor and VS Code it is their MCP or connector settings. The client opens the UserLimit sign-in, you approve access, and the tools are ready.

Can the AI start a load test without asking me?

No. Before any traffic is sent, the server returns a summary of the run: the site, the number of users, the duration and the hours it will use. The assistant has to show it to you and get your yes. The server only accepts a start that matches that exact summary, within 15 minutes.

Can the AI spend money or buy hours?

No. The assistant can see how many hours you have left and estimate what a run will cost, but buying hours, plans and auto top-up stay on userlimit.com. When you run short, it gives you a link to the billing page.

Which sites can I test from an AI assistant?

Only domains you have verified with a DNS TXT record or a meta tag. The assistant can add a domain and tell you which record to create, but the check is the same as in the web app, so nobody can point a test at a site they don't own.

How does the AI write the test scenario?

The assistant writes a short Playwright script from your description, such as open the homepage, click pricing, filter by T-shirts. UserLimit then sends one free session through it in a real browser. If a step fails, it gets the step and the error and fixes it before you run anything.

Which AI assistants does it work with?

Any MCP client that supports remote servers with OAuth 2.1 and Dynamic Client Registration: Claude (web, desktop, mobile and Claude Code), ChatGPT, Cursor, VS Code, Windsurf and other agents. They all connect to the same URL.

How is it different from a k6 or JMeter MCP server?

Those servers drive a load tool you install and run yourself, and they send protocol-level requests. UserLimit is hosted: the sessions are real browsers in the cloud, so the test measures what visitors see, including page loads and failed clicks. There is nothing to install, and the domain checks, confirmations and limits are enforced by the server.

Is there a free tier?

Yes. A new account gets 3 free virtual-user hours, enough for the free scenario check and a small first test, with no card. Larger tests use pay-as-you-go hours or a plan.

Every site has a limit

Your next load test is one message away

3 free hours to start, no card, nothing to install.