AI test automation over MCP, from the assistant you already use

Works with Claude Code · Cursor · VS Code · Claude Desktop

What happens, in four steps

1. ConnectCreate a key in Settings and paste the setup line into your assistant.
2. Ask“Why did Checkout fail?” or “write a test for this story.”
3. It checksIt reads the runs and counts each candidate locator on the live page before proposing anything.
4. It recordsA change is a version in the test’s history under your name, and the run it starts is an ordinary run in Reports.

Powerful enough to maintain tests, constrained enough to trust

An assistant decides which calls to make, partly from text written by whoever controls the application under test: page content, error messages, step names. So the limits are enforced by QAright’s API, where a model cannot talk its way round them, rather than by the assistant’s good behaviour.

What the assistant can reach

QAright serves an MCP server at https://api.qaright.com/mcp. An assistant connected to it gets tools for the work around a browser suite, not a browser of its own:

• Suites and tests. List suites, read a suite’s tree and a test’s steps, create manual and automated tests, add a step, link a test to a requirement.

• Runs. Run a test or a suite, rerun a case, and read the report, the history and the pass-rate trend.

• Failures. A case’s failure history, the difference from its last passing run, a diagnosis of a failed run, the screenshot of a failed step, and which tests are flaky.

• Locators. The locators recorded for a page, the ones that keep failing, a live check of a locator before anything is saved, and a proposed fix for a locator or a wait.

• Context. The pages and elements QAright has learned about your app, and the existing tests that already cover something similar to a new story.

Three jobs it is good at

• Fixing a test that broke overnight.The assistant reads which step failed and how often, compares the run with the last one that passed, and counts each candidate locator on the live page, ranked unique and visible first. The chosen one is applied as a single step update and pinned, so a later automatic heal does not quietly undo it.
• Writing a test from a story.Given a user story, it looks for existing tests that already cover part of it, then writes the steps and counts each locator on the opening page before saving. An unreadable locator refuses the save and one that matches several elements is saved only when explicitly allowed; one absent from the opening page is fine, because it may belong to a later page, and the result says which.
• Finding what is flaky.It asks which tests pass and fail without anything changing, classifies the recent failures, and proposes a wait where the evidence says the page was slow rather than broken. The difference is the one flaky tests explains.

Connecting one

On Solo and above, open Settings in QAright, choose AI assistant under Automation, and create a key. The page then shows the setup for each assistant with the key already in it. For Claude Code it is one command:

claude mcp add --transport http qaright https://api.qaright.com/mcp \
  --header "X-API-Key: <your companion key>"

VS Code asks for the key the first time rather than keeping it in a workspace file that could be committed, and Claude Desktop connects through mcp-remote. Make one key per assistant, so that revoking one leaves the others working.

How this differs from Playwright MCP

Microsoft’s Playwright MCP server gives a model a browser: it can open a page, read it and click. That is useful for exploring an app. QAright’s server gives a model your test suite instead: the runs, the failure history, the recorded locators and their fallbacks, the healing that already happened, and a live check before a change is saved. The first answers “what is on this page?”; the second answers “why did this test fail, and what is the smallest safe fix?”

What it costs, and what it does not do

It uses none of your plan’s AI allowance. The reasoning is your assistant’s own; QAright runs the tests and answers its questions. Runs it starts happen on QAright’s servers, so they reach public addresses, not localhost or an intranet. And it does not record: recording still happens in your browser, through the extension or the desktop agent.