The usual way
- Launches a brand-new, empty browser
- You script logins and seed sessions first
- Feature flags and SSO have to be recreated
- Tests run as a bot, not as you
- Brittle selectors break on every change
AI Testing · MCP Server + Browser ExtensionFree
Your AI agent tests in the browser you're already signed into.
AutoQAI connects Claude Code, Cursor, GitHub Copilot, or Kiro to the exact tab you have open. Cookies, SSO, and feature flags stay live — so the agent runs real QA flows, assertions, and visual comparisons against your actual app, and hands you findings with the proof attached.
- No fresh browser
- No re-login
- Loopback-only · private
- Works in your editor
Quick start
# 1 · Install the MCP server
npm install -g @autoqai/mcp
# 2 · Register it with your AI agent
autoqai init
# 3 · Start it inside your project
autoqaiWorks with
Claude CodeCursorGitHub CopilotKiro
The difference
Headless tools start from nothing. AutoQAI starts from your real session.
With AutoQAI
- Drives the tab you already have open
- Cookies, SSO, and tokens stay live
- Feature flags evaluate in your context
- Tests the app exactly as you experience it
- The agent adapts instead of replaying a script
How it connects
Three pieces, one private connection
The agent never launches a browser. It talks to a small local server, which relays instructions to the extension running in your tab — all over a loopback link that never leaves your machine.
AI CLI Agent
Claude Code, Cursor, GitHub Copilot, or Kiro. It talks to AutoQAI over the Model Context Protocol and decides what to test and how.
AutoQAI MCP Server
A small Node process started inside your project. It exposes testing capabilities to the agent and bridges them to your browser over a private loopback connection.
Browser Extension
Pairs with the server and runs each instruction against the tab you bound — your real, authenticated session — using the browser's own DevTools protocol.
See it work
One loop, from your tab to a verified fix
Bind the tab you're already in, queue what you want checked, and your AI agent drives the rest — testing, capturing proof, and verifying fixes in your real session. Watch the loop run, or step through it yourself.
01
Bind your tab
Pair the extension and attach the tab you're already signed into. Cookies, SSO, and feature flags stay live.
Capabilities
What the agent can do on the page
AutoQAI gives the agent a full toolkit for inspecting, operating, verifying, and comparing your application — the same things a careful QA engineer would do.
Interact like a person
The agent operates the page through real user-style actions, not brittle scripts.
- Click, hover, and drag-and-drop between elements
- Type into fields, fill inputs atomically, and press individual keys
- Choose options from dropdowns and custom selects
- Upload files into file inputs
- Wait for elements to appear, conditions to become true, or the network to go idle
Understand the page
Before acting, the agent can read the structure and state of what's on screen.
- Inspect the DOM and accessibility tree
- Read page metadata — URL, title, viewport, and scroll state
- Pull structured data out of tables and form fields, including ARIA custom controls
- Capture screenshots of the page or a specific region
- Run read-only checks against the live page
Assert behaviour
Verifications that turn into findings automatically when they fail.
- Assert visible text, element visibility, and enabled / disabled state
- Check input values, checkbox / radio / switch state, and CSS classes
- Assert element counts with comparison operators
- Verify computed styles match expectations
- Confirm network requests succeeded and the console is free of errors
Compare old vs new
Deterministic, offline diffing for redesigns and regression sweeps — no tokens spent reasoning over raw data.
- Extract and diff data tables, resolving merged cells and per-cell formatting
- Compare forms field-by-field — labels, values, types, required state, and help text
- Diff charts by series and data point (Highcharts, Chart.js)
- Compare metric cards and labelled-value sections
- Compare computed styles property-by-property
Plan, run, and report
A structured testing loop that produces reviewable results.
- Draft a plan of test cases for each testing task
- Track per-test-case status: passed, failed, or skipped
- Record findings — observations, warnings, and confirmed bugs
- Reproduce bug reports captured straight from the page
- Finalize each run with a summary you can review in the dashboard
Build a regression library
Curate reusable test cases once, then re-run them after every change — organised the way your product actually works.
- Organise test cases as Modules → Features → cases, version-controlled in your repo
- Pick stable selectors straight from the live page instead of typing them
- Let the AI draft a starter library from your project context — or from the live screen
- Queue any mix of whole modules, features, or individual cases for a run
- Get per-case passed / failed / skipped status, grouped back into each run
Work with your tools
AutoQAI fits into the workflow you already have.
- Pull Jira ticket details through your authenticated session
- Attach linked tickets to testing tasks and bug reports
- Keep a project context file the agent reads before every run
- Capture console and network trails alongside each step
- Drive multiple tabs in parallel for multi-window or A/B flows
Look & feel
See it in action
Private by design
Nothing phones home
The server and extension talk only over a loopback connection on your own machine, secured with a pairing token. Connections from anywhere else are refused. Your runs, findings, and screenshots are written as plain files inside your project — you own every byte.
- Loopback only. Bound to 127.0.0.1; non-local connections are rejected at connect time.
- Token-paired. The extension only attaches after you enter the pairing token shown by the server.
- Your session, untouched. Credentials and cookies stay in your browser; AutoQAI never collects or transmits them.
- Local artifacts. Everything lands in a
.autoqaifolder in your repo — no cloud, no telemetry.
Testimonials
What teams say
"The killer feature is that it tests in my logged-in session. No more writing throwaway auth setup for every flow — I just point my agent at the tab and it goes."
"The deterministic table and form comparisons are brilliant. We caught a dozen subtle regressions during a redesign that a human eyeballing two tabs would never have spotted."
"Pairing took thirty seconds and from then on my AI agent could click through real flows and hand me findings with screenshots attached. It feels like having a tireless QA partner."
"Everything stays on localhost and writes to files in the repo. That made it an easy yes for our security review — there's nothing phoning home."
FAQ
Frequently asked questions
How is this different from Playwright or a headless browser?
Headless tools launch a brand-new, empty browser — so you have to script logins, seed sessions, and recreate state before any real test can run. AutoQAI drives the browser you already have open and signed in. Your cookies, SSO, and feature flags are live, so the agent tests the application exactly as it behaves for a real, authenticated user.
Which AI agents does it work with?
AutoQAI works with any MCP-capable CLI agent. It can auto-register itself with Claude Code, Cursor, GitHub Copilot (agent mode in VS Code), and Kiro using a single init command, and it works with any other client that supports MCP servers over stdio.
Does my data or my page content leave my machine?
No. The MCP server and the extension communicate over a loopback-only WebSocket (127.0.0.1) protected by a pairing token. Connections from any non-loopback address are refused. Test runs, findings, and screenshots are written to plain files inside your project folder.
Can the agent control any tab without my permission?
No. You explicitly pair a project and bind a specific tab from the extension. The browser shows its standard debugging indicator while a tab is bound, and you can unbind at any time.
What can the agent actually do on the page?
It can inspect the DOM and accessibility tree, click, type, fill, hover, drag, select options, upload files, press keys, wait for elements or network idle, run assertions, capture screenshots, and extract and compare structured content like tables, forms, and charts.
Is it really free?
Yes. AutoQAI is free to use. The MCP server is published on npm and the extension is available from the Chrome Web Store.
Can it test more than one tab at once?
Yes. Each bound tab gets its own connection, so an agent can drive several tabs in parallel — useful for multi-window flows or comparing two environments side by side.
Where are my test runs and findings stored?
Inside your project, under a local .autoqai folder. Configuration, the project context file, the task queue, and every run's test cases, steps, findings, and screenshots are kept there as JSON, Markdown, and image files you fully own.
Can I save reusable test cases for regression testing?
Yes. The dashboard's Test cases section lets you build a regression library organised as Modules, Features, and individual cases — each saved as a tracked file in your repo. You can author cases by hand (picking selectors straight from the page) or have the agent draft a starter library, then queue any selection for the agent to run, getting a passed / failed / skipped result per case.
How do I remove AutoQAI from my agent?
Run autoqai uninstall to reverse the init command — it removes only the autoqai entry from each client's MCP config and leaves everything else untouched. Target specific clients with --client, scope it to a project with --project, or preview with --dry-run. Your project's .autoqai folder is never touched, so delete that by hand if you also want to remove a project's runs and queue.
Is there an easy way to set up without remembering commands?
Yes. Run autoqai with no arguments in a terminal and it opens a guided menu that walks you through registering it with your AI client, pairing the extension, and listing your projects. When your agent launches the same command (without an interactive terminal) it runs the MCP server instead, so the integration is never affected.
After I pair, why does the server keep running?
Pairing is a live connection, so the pairing flow keeps the server running to keep the extension connected — and lets you pair more tabs. When you're ready to use the AI features, stop it (or press Ctrl-C) and start your agent, which launches its own AutoQAI server on the same project port so the bound tab keeps working.
Do I need to restart my IDE after setup?
If your AI client was already running when you registered AutoQAI, restart it so it loads the new MCP server. For IDE-integrated extensions (VS Code, Cursor, Kiro) that usually means reloading or restarting the IDE window itself, not just the chat panel. A standalone CLI like Claude Code just needs to be relaunched in the project.
Point your agent at a real browser tab.
Install the package, add the extension, and pair in under a minute. The full walkthrough is in the docs.