Let Cursor check its own work in a real browser
Connect Testsy to Cursor over MCP so the agent can save a plain-language test, run it against your deployed app or a phone build, and read the verdict.

Why the agent should not grade itself#
Cursor can write a feature and a test for it in the same breath. Both come from the same context, so they share the same assumptions: if the agent misunderstood what "save the draft" means, the test checks the misunderstanding. When the tests pass, they tell you the code matches what the agent believed.
A check from outside does not share those beliefs. It opens your deployed app, does what the sentence says, and reports what was on the screen. That is the job Testsy does, and over MCP Cursor can ask for it directly.
Connect it#
Two ways, from the MCP docs:
- The plugin. testsy-ai/testsy-cursor holds the MCP configuration and three skills.
- By hand. Create an API key in Settings → API keys (choose the kind that can also create and edit tests if Cursor should save new tests), store it as
TESTSY_AI_API_KEY, and add the server to Cursor's MCP settings:
{
"mcpServers": {
"testsy": {
"url": "https://app.testsy.dev/mcp",
"headers": { "Authorization": "Bearer qa_…" }
}
}
}Keep the key in an environment variable or secret store. Do not paste it into a prompt or commit it. To get the skills without the plugin, copy the skills/ folders from the plugin repository into your project's .cursor/skills/.
The three skills are testsy-check-release (check a change before you call it done), testsy-investigate-run (read a failed result and find the cause) and testsy-regression-from-bug (turn a bug into a saved test).
What Cursor does with it#
Ask in plain words: "Check that a new user can sign up and create a project on our staging site." The agent works through the same steps you would in the app:
list_projects, thenplan_testwith what you asked. Testsy drafts the steps and the checks.- It shows you the plan. You correct it ("the button is called Create workspace, not New project") and it calls
plan_testagain with the change. create_testsaves the confirmed plan;run_teststarts it and returns a watch link you can open to see the run live, no sign-in needed.wait_rununtil the verdict is final, then it reads each check, its reason and the value it was decided on.- If a check failed, it fixes the code, deploys, and runs the same test again.

The test is saved in your workspace. Next week, after a different change, Cursor (or you, or CI) runs it again.
Apps too#
For an iPhone or Android app, the agent can upload the build it just made and run the app's tests on it: find_build to reuse a build that is already there, otherwise create_build_upload and a shell command that sends the file. The steps are in the MCP docs.
Keep a person in the loop#
- Review the plan before the first run. Runs can create accounts and change data, so point them at staging with made-up details.
- Cursor's own approval settings decide when it asks before a tool runs. Leave approvals on until you trust the flow.
- Read the evidence, not only the verdict. Every result keeps the screen each check was read from.
Every run counts toward your plan like any other. The free trial has 10.
This post is part of Testing an app you built with AI. The same setup for other agents: Claude Code, Codex, Gemini CLI and the Claude app.
Filed underWebAI-built appsCI, schedules and agents