The Best Free Alternative to Manual Playwright: 4 MCP Tools That Replace Your .spec.ts Files

1. Why "Which Framework?" Is the Wrong Question

Search for "best Playwright alternative" and you get lists of 8 to 12 testing frameworks. Cypress, Selenium, TestCafe, Puppeteer, WebdriverIO, Katalon. Each article compares syntax, language support, and pricing. They assume you want to keep writing test code but in a different tool.

That assumption misses the actual problem. The pain of manual Playwright is not Playwright itself. It is maintaining hundreds of .spec.ts files, updating selectors when the UI changes, debugging flaky waits, and spending engineering hours on test infrastructure instead of product work.

Switching from Playwright to Cypress does not fix that. You still write code. You still maintain selectors. You still debug flaky tests. The real alternative is removing the test code layer entirely and letting an AI agent handle browser interaction from plain English instructions.

2. The 4 MCP Tools in server.ts

Assrt is an MCP server. MCP (Model Context Protocol) is the standard that lets AI coding agents call external tools. When you install Assrt, it registers 4 tools that your agent (Claude Code, Cursor, or any MCP-compatible client) can call directly:

These 4 tools are defined in src/mcp/server.ts of the Assrt MCP package (894 lines). That one file is the entire interface between your AI agent and browser testing. No test runner binary, no config files, no plugin system.

3. What Happens When You Run assrt_test

When your AI agent calls assrt_test with a URL and a test plan, here is what happens under the hood:

1. A Freestyle cloud VM spins up. Not a Docker container on your machine. An ephemeral VM at a unique *.vm.freestyle.sh domain. First boot takes about 11 seconds. If a snapshot exists for this configuration, the VM restores from snapshot with all services pre-warmed.
2. A real Chromium browser opens inside the VM. The agent controls it via Playwright MCP, using tools like snapshot, click, type_text, scroll, press_key, wait, and assert. These are real Playwright actions, not simulated clicks.
3. The AI agent (Claude Haiku or Gemini) interprets each test case. It reads the plain English step, decides which browser tool to call, executes it, observes the result via DOM snapshot, and moves to the next step. No hardcoded selectors. The agent finds elements by understanding the page.
4. Results stream back in real time. Your IDE receives screencast frames (~15fps), step completions, assertion results, and UX improvement suggestions as they happen. A video is recorded for post-run review.
5. Structured results are saved locally. The test plan goes to /tmp/assrt/scenario.md and results to /tmp/assrt/results/latest.json. Every run gets a unique scenario ID (UUID) so you can re-run the exact same test later.

The entire flow happens without you opening a terminal, writing a test file, or configuring a test runner. You describe what to test in your IDE chat. The agent handles everything else.

4. Manual Playwright vs. Assrt: Same Signup Flow

Here is the same test written both ways. A user signs up, verifies their email, and lands on the dashboard.

Manual Playwright (TypeScript)

import { test, expect } from '@playwright/test';
import Mailosaur from 'mailosaur';

const mailosaur = new Mailosaur(process.env.MAILOSAUR_KEY!);

test('signup flow', async ({ page }) => {
  const email = `test.${Date.now()}@srv.mailosaur.net`;
  await page.goto('https://myapp.com/signup');
  await page.getByLabel('Email').fill(email);
  await page.getByLabel('Password').fill('Pass123!');
  await page.getByRole('button', { name: 'Sign Up' }).click();

  const msg = await mailosaur.messages.get('srv', {
    sentTo: email
  }, { timeout: 30000 });

  const code = msg.html.body.match(/\d{6}/)?.[0] ?? '';
  await page.getByLabel('Code').fill(code);
  await page.getByRole('button', { name: 'Verify' }).click();
  await expect(page.getByText('Dashboard')).toBeVisible();
});

This requires a Mailosaur subscription ($99+/month), an API key in CI secrets, regex parsing for the OTP code, and manual timeout handling. When your email template changes, the regex breaks silently.

Assrt (Plain English)

#Case 1: Full signup with email verification
1. Create a temporary email address
2. Navigate to the signup page
3. Type the temp email into the email field
4. Type a password into the password field
5. Click Sign Up
6. Wait for the verification code in the email
7. Type the verification code into the OTP input
8. Click Verify
9. Verify the dashboard loads

5. No Vendor Lock-In: Your Tests Are Markdown

Most AI testing tools trap your tests in a proprietary format or a cloud dashboard you cannot export from. When you stop paying, your tests disappear.

Assrt tests are markdown files. The #Case N: format is plain text that any human can read. After each run, the plan is saved to /tmp/assrt/scenario.md on your local filesystem. Results go to /tmp/assrt/results/latest.json as structured JSON with pass counts, fail counts, duration, and per-scenario assertions with evidence strings.

The npm package (@m13v/assrt) is MIT licensed. The Dockerfile in the repo lets you self-host on your own infrastructure. You can run the MCP server locally, point it at your own LLM provider, and keep every piece of the stack under your control.

If you stop using Assrt tomorrow, you keep your markdown test plans, your JSON results, and your video recordings. Nothing is locked behind a login or an API key.

6. Frequently Asked Questions