Generate tests with AI

Describe a flow in plain language, and the agent writes and runs the test.

Describe a flow in plain language, and the TestDriver agent writes the test for you.

• The agent runs in your AI coding assistant.
• It runs the flow in a live sandbox and writes a Vitest test.
• It finds elements by looking at the screen, so you never write selectors.

Before you begin

Install TestDriver in an AI coding assistant that can edit files in your repository:

• Claude Code: install the plugin.
• VS Code or Cursor: install the extension.
• Any other assistant: run npx testdriverai init in your repository and choose your assistant. See Set up a project.

Start the agent

Ask for the TestDriver agent by name in your assistant’s chat:

Use the testdriver agent to write a test that logs in to https://testdriver-sandbox.vercel.app/login

In VS Code, you can also pick testdriver from the agent dropdown in the Copilot chat panel.

If your assistant doesn’t support agent files, give it the TestDriver agent instructions directly.

Describe the test

Write the flow as the steps a user takes and what they should see at the end.

Explicit steps make a test faster, cheaper, and more reliable than one where the agent works out the flow itself:

Write a TestDriver test that does the following:

1. Open https://practicetestautomation.com/practice-test-login/
2. Type "student" into the Username field
3. Type "Password123" into the Password field
4. Click the Submit button
5. Verify the new page says "Logged In Successfully"

Watch and steer the run

The agent sends a screenshot after every action, so you see what it sees.

• In clients that support MCP Apps, each screenshot appears inline in the chat, with the targeted element highlighted.
• With the TestDriver extension in VS Code or Cursor, the sandbox also plays live in a panel beside your code.

If the agent gets stuck or takes a wrong turn, tell it what to do next:

A cookie banner is covering the form. Accept it, then continue.

You can also ask for more checks at any point:

Also assert that the page shows a Log out button.

What the agent does

1. Starts a session. It provisions a sandbox (Linux by default) and opens your app, for example Chrome at the URL you gave it.
2. Performs each step. It finds each element from a description, acts on it, and checks the screenshot to confirm the step worked.
3. Writes code as it goes. After each successful action, it appends the matching SDK code to a test file under tests/. The file always reflects what has actually run.
4. Runs the test. It runs the file with vitest run and keeps fixing it until it passes.

Review the generated test

The result is an ordinary Vitest file. Read, edit, commit, and run it in CI like any other test:

tests/login.test.mjs

import { describe, expect, it } from 'vitest';
import { TestDriver } from 'testdriverai/vitest/hooks';

describe('Practice login', () => {
  it('logs in with valid credentials', async (context) => {
    const testdriver = TestDriver(context);

    await testdriver.provision.chrome({
      url: 'https://practicetestautomation.com/practice-test-login/',
    });

    const usernameField = await testdriver.find('Username input field');
    await usernameField.click();
    await testdriver.type('student');

    const passwordField = await testdriver.find('Password input field');
    await passwordField.click();
    await testdriver.type('Password123', { secret: true });

    const submitButton = await testdriver.find('Submit button');
    await submitButton.click();

    const result = await testdriver.assert('the page says "Logged In Successfully"');
    expect(result).toBe(true);
  });
});

Before you commit, check two things:

• Each find() description would still pick the right element if the page changed.
• The assertions cover what you care about.

If the test calls ai(), which works out a step on its own at run time, ask the agent to replace it with explicit find(), click(), and type() steps before you commit. Explicit steps run faster and do the same thing every time.

Tips for better tests

Describe elements by what they look like

“The blue Sign In button in the header” works better than “the button”. Include the visible label. Add position or color when several elements look alike. See Locate elements.

Mention loading and animations

If part of your app loads slowly, say so. For example: “wait for the loading spinner to disappear before continuing”. The agent then waits instead of acting too early.

Assert after each important step

Assert after each meaningful action, such as “the item appears in the cart”. When the test fails later, you see where the regression starts.

Build long flows in pieces

Don’t describe a long workflow all at once. Have the agent get the first part passing, then describe the next.

Next steps

Run tests

Choose a device, preview mode, and how to run your tests.

Speed up reruns with caching

Reuse what TestDriver found so reruns skip the AI.

Make assertions

Write checks that catch the regressions you care about.

Run in CI/CD

Run generated tests on every pull request.