Let AI Control Your Real Browser — Not a Throwaway One

Most browser MCP servers launch a separate browser instance. Clean slate, no cookies, no auth. Want to automate something behind a login? Re-authenticate every time. Want to test your actual session state? Too bad.

What if the AI could just use the browser you already have open?

That's what Dramaturg + @playwright-repl/mcp does. The AI controls your real Chrome tabs — your cookies, your sessions, your localStorage — all intact.

How it works

Claude (or any MCP client)
  ↕ MCP (stdio)
playwright-repl MCP server
  ↕ WebSocket
Dramaturg Chrome extension
  ↕ CDP / chrome.debugger
Your real Chrome tabs

The key difference: Playwright runs inside the browser, not outside it. The extension uses playwright-crx to run the full Playwright API directly in Chrome's service worker. No Node.js relay, no separate browser process.

This means:

• No re-authentication — Gmail, Notion, your internal tools — already logged in
• expect() assertions — something no other browser MCP server supports
• You see everything — the AI works in your actual tabs, in real time

Just two tools

While Playwright's official MCP server exposes ~70 tools, @playwright-repl/mcp exposes just two:

• run_command — execute a single command (keyword or JavaScript)
• run_script — run a multi-line script

The AI figures out what to do with them. Here's what a command looks like:

Keyword syntax (.pw)

snapshot                    → accessibility tree
goto https://example.com    → navigate
click "Sign in"             → click by text
fill "Email" "user@test.com"  → fill a form field
verify-text "Welcome"       → assert text is visible

Playwright API (JavaScript)

await page.getByRole('button', { name: 'Sign in' }).click();
await expect(page.getByText('Welcome')).toBeVisible();

Both work through the same run_command tool. The extension auto-detects the format.

AI Agents that actually verify their work

The MCP package includes four agents that go beyond "generate and hope":

Planner — Give it a URL and a goal. It takes snapshots, maps out the page structure, and produces a step-by-step workflow plan with the exact labels and text it found.

Generator — Takes a plan and turns it into a working script. The key word is working — it runs every command in the real browser, assembles the script, executes the full thing via run_script, and iterates until it passes.

Healer — Give it a broken script. Wrong selectors, changed page structure, timing issues. It runs the script, diagnoses failures with snapshots, fixes them, and re-runs until all lines pass.

Converter — Translates between .pw keyword syntax and JavaScript Playwright API. Produces idiomatic output — proper locator strategies, extracted variables, chained methods.

Every agent must run the script before outputting it. No "here's what should work" — it either passes or the agent keeps fixing.

Setup

1. Install

npm install -g @playwright-repl/mcp

Install Dramaturg from the Chrome Web Store.

2. Configure Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "playwright-repl": {
      "command": "playwright-repl-mcp"
    }
  }
}

Or for Claude Code:

claude mcp add playwright-repl playwright-repl-mcp

3. Install the AI agents (optional)

mkdir -p .claude/agents
cp node_modules/@playwright-repl/mcp/agents/*.agent.md .claude/agents/

4. That's it

The extension connects automatically. No side panel needed. Just make sure Chrome is running.

Example prompts

Go to https://demo.playwright.dev/todomvc, add three todos, complete one, and verify the Active filter only shows the remaining two.

@playwright-repl-planner explore https://demo.playwright.dev/todomvc and plan a test for adding and completing todos

@playwright-repl-generator create a .pw script that adds 3 todos, completes one, and verifies the Active filter

@playwright-repl-healer this script is failing on line 5 — the selector changed after a site update: [paste script]

@playwright-repl-converter convert this .pw script to JavaScript: [paste script]

Human + AI, side by side

The extension isn't just a bridge for the AI — it's a full REPL you can use alongside it:

• Watch commands execute in real time as the AI works
• Intervene manually — type snapshot to check state, or navigate to a different page
• Record your own interactions — click Record, interact with the page, get .pw commands or JS code
• Switch tabs — the toolbar dropdown lets you redirect both the AI and your commands to any open tab

vs. other browser MCP servers

Try it

> snapshot
> click "Sign in"
> fill "Email" "user@test.com"
> verify-text "Dashboard"

Your browser. Your sessions. AI that verifies its own work.

• Chrome Web Store — Dramaturg
• GitHub — playwright-repl
• npm — @playwright-repl/mcp