AutoQAI

Let your AI coding agent test in your real, logged-in browser tab.

AutoQAI connects MCP-capable AI CLI agents — Claude Code, Cursor, GitHub Copilot, Kiro, and others — to the browser tab you already have open and signed in. Instead of spinning up a fresh, unauthenticated Selenium or Playwright session that has no idea who you are, your agent drives your real page: clicking, typing, filling forms, taking screenshots, reading the DOM, and checking the console and network.

Cookies, SSO, MFA, tenant, and feature flags stay exactly as they are, because nothing new is launched. When the agent says "open the dashboard and click Save", it runs inside the same browser you are testing with — same session, same tenant, same flags.

📖 Full setup guide, supported agents, and documentation: https://utilsware.com/products/autoqai

⚠️ This package needs the browser extension

AutoQAI has two halves. This npm package is the MCP server; it does nothing without the companion browser extension:

1. MCP server (this package) — your AI agent talks to it.
2. Browser extension — drives your tab. 👉 Install it from https://utilsware.com/products/autoqai

The agent talks to the MCP server, the server talks to the extension over a local-only (loopback 127.0.0.1) WebSocket, and the extension drives your tab. Without both installed and paired, there is nothing to connect to.

What your agent can do in your tab

• Interact — click, fill, type, press keys, select options, hover, drag, and upload files.
• Observe — read console logs, network activity, and page issues to diagnose real bugs.
• Capture — screenshots of the full page or specific elements/regions.
• Compare — old vs new, before vs after, tab vs tab: tables, forms, charts, sections, and styles.
• Verify — assert text, values, visibility, counts, styles, network success, and "no console errors".
• Work across multiple tabs in parallel, each with a short label (e.g. main, qa-tenant, mobile).

A typical use: ask your agent to reproduce a bug in your staging environment, fix the code, and confirm the fix — all inside the session you are already logged in to.

Who it's for

• Developers who want their AI agent to reproduce and fix bugs inside their own browser session.
• QA engineers who want live console/network observability and screenshot-based bug capture with a one-click "ask the agent to fix this".
• Anyone who has hit the wall of "the AI's test browser isn't logged in to my staging environment."

Installation

1. Install the MCP server (this package)

npm install -g @autoqai/mcp

Requires Node.js 22+. This adds the autoqai command to your PATH.

On an older extension? Web Store reviews can take a couple of days, so a brand-new MCP release on npm may briefly run ahead of the extension version available in the store. If the two halves disagree, install the MCP version that matches the extension you currently have until the new extension is approved — for example:

npm install -g @autoqai/[email protected]

Once Chrome ships the new extension, switch back to the latest with npm install -g @autoqai/mcp (or @latest).

2. Install the browser extension

Get it from https://utilsware.com/products/autoqai and pin it.

3. Register AutoQAI with your AI agent

On a clean machine, installing the package registers AutoQAI for you — it writes the MCP registration into every supported agent at user scope the first time it's installed. It only runs when no AutoQAI registration exists yet, so updating an existing install (or one where you removed a client on purpose) leaves your configuration untouched.

If the automatic step didn't run — because you already had AutoQAI configured, or the install happened in CI — register it yourself with one command that covers every supported agent it finds installed:

autoqai init

Then restart your AI agent so it picks up the new server. To register a single agent, add --client <name> — one of claude-code, cursor, vscode, copilot, kiro, or visualstudio (for example --client claude-code).

Using GitHub Copilot? GitHub Copilot's agent mode in VS Code reads VS Code's mcp.json, so --client copilot and --client vscode register the same file — and a plain autoqai init already covers it. After registering, switch Copilot Chat to Agent mode, open the Tools picker, and make sure autoqai is enabled.

Using a JetBrains IDE (IntelliJ, PyCharm, Rider, WebStorm)? JetBrains AI Assistant configures MCP servers from its Settings UI rather than a file on disk, so there's no init command for it. Open Settings → Tools → AI Assistant → Model Context Protocol (MCP), add a new stdio server named autoqai with the command autoqai, and apply (or use Import from Claude). Requires AI Assistant with MCP support (IDE 2025.1+).

Claude Code — ~/.claude.json:

{
    "mcpServers": {
        "autoqai": { "command": "autoqai" },
    },
}

VS Code / GitHub Copilot — <user-profile>/Code/User/mcp.json:

{
    "servers": {
        "autoqai": { "type": "stdio", "command": "autoqai" },
    },
}

Visual Studio — ~/.mcp.json (or ./.mcp.json next to your solution):

{
    "servers": {
        "autoqai": { "type": "stdio", "command": "autoqai" },
    },
}

Getting started

1. Start AutoQAI in your project. Your AI agent runs the AutoQAI server from your project directory. It prints a host, a port, and a short pairing token.

2. Pair the extension. Open the extension popup. Your running project appears under Discovered projects with its host and port already filled in — click it and paste the pairing token once (the popover remembers it afterwards). If it isn't listed, click Rescan, or choose Add a project manually and enter the host (127.0.0.1), port, and token, then click Pair & connect.

3. Bind a tab. Choose the tab you want the agent to control, give it a short label (e.g. main), and attach. Chrome shows its standard "is being debugged" banner — leave it visible while the agent works. While the agent is driving a bound tab, any new tab an action opens (a link in a new tab, a popup, an OAuth screen) is bound to the same project automatically, so the agent keeps going without you binding it by hand.

4. Ask your agent to test. Tell it to test a flow, reproduce a bug, or compare two screens. It drives your real tab and reports findings back. You never call any low-level commands yourself.

CLI reference

| Command / option | Description | | ----------------- | ------------------------------------------------------------------------- | | autoqai | Start the MCP server for the current project directory. | | autoqai init | Register AutoQAI with supported AI agents (--client <name> for one). | | --port <n> | Pin the server port (default 7820; auto-picks a nearby port if in use). | | --config <path> | Use a specific config path. | | autoqai --help | Show all options. |

Privacy & security

• Local-only. The extension accepts connections only from 127.0.0.1 (loopback). Non-local connections are refused — no LAN exposure.
• Per-project pairing token generated by the server and entered once.
• Debugger attaches only after you bind a tab, and Chrome's "is being debugged" banner stays visible the whole time.
• No credentials, OAuth tokens, or cookies are stored or transmitted by the extension.
• No analytics, no telemetry, no third-party SDKs. Page content (DOM, screenshots, logs) only leaves your browser when the agent you started explicitly requests it.
• The agent acts inside your existing session — it never launches a separate browser or asks for your passwords.

Troubleshooting

• Pairing fails → make sure the extension is connecting to 127.0.0.1 (loopback only).
• autoqai not found after global install → restart your terminal so the npm global bin is on your PATH (check with npm config get prefix).
• Chrome refuses to attach → bind a regular http(s) tab; Chrome blocks attaching to privileged pages (e.g. chrome://, the Web Store).

Links

• 📖 Documentation & extension download: https://utilsware.com/products/autoqai/docs
• 📄 License: https://utilsware.com/products/autoqai/license

License

See the license terms.