QAJev MCP Server

Test your website, game or app the way a person uses it. Say what a visitor wants in plain words; Jev, a small, fast model, clicks through a real Chrome to do it (or plays your game, or taps through your iOS or Android app); QAJev judges the result from the page itself and gives you a clear report. From the command line, or from any AI agent through MCP. qajev.com

qajev check https://shop.example \
  --goal "Find what the Pro plan costs per month. Stop when that price is visible." \
  --expect-text '$29 per month'

Why QAJev

• Plain-language tests. No selectors to maintain. "Find the refund policy. Stop when the number of days is visible." still works after a redesign.

• Verdicts you can trust. Jev saying "done" is never proof: your expectations decide. And a problem with the product (fail) is never mixed up with a problem on the tool's side (harness) or a visitor getting lost (stuck, often a real usability issue).

• Desktop and phone, always. Every website test also runs in a phone view (a phone's screen size, touch and user agent, like a browser's device mode), unless you pin one device. More.

• Fast and cheap. A decision takes about a quarter of a second and costs a fraction of a cent. The smoke crawl is free.

• Safe by default. Production is read-only and destructive is never: a run that would change a site that is not on your machine is refused, and delete and refund buttons are hidden. Jev never types passwords or payment details, and QAJev types a password only for a seeded test account on your own machine (localhost, *.test); for any other account you sign in once in QAJev's window (qajev browser login) and it keeps that session. A hard cost cap on every run. More.

• Built for agents. An MCP server, a ready-made agent prompt, jobs that any agent can follow or stop, and one shared queue so many agents can share a machine.

• See it live. qajev top in a terminal, or qajev dashboard in the browser: every run, what Jev is doing right now, what it costs, and each test with its screenshots.

• Games too. Godot and Electron games: Jev works the menus, the game's own bot plays in real time, QAJev watches the frame rate, memory and the game's state.

• Mobile too. Native apps and mobile websites on the iOS Simulator and Android emulator, on a throwaway device QAJev starts and removes itself. More.

Related MCP server: pursr-mcp

Quick start

uv tool install git+https://github.com/hanamorilabs/qajev     # or: pipx install git+https://github.com/hanamorilabs/qajev
mkdir -p ~/.qajev && echo 'OPENROUTER_API_KEY=sk-or-...' >> ~/.qajev/.env   # or TYPESAFE_API_KEY
qajev doctor                                                # checks your setup

qajev smoke http://localhost:3000                           # free: errors, broken links, missing basics
qajev check http://localhost:3000 --goal "Open the pricing page. Stop when the prices are visible." \
  --expect-url /pricing

Every run writes report.html (open it in a browser), report.md and report.json. Details: Getting started.

What you can do

Use it from an AI agent

claude mcp add -s user qajev -- qajev mcp       # Claude Code; see docs/mcp.md for Codex, Cursor and others

Or just tell your agent to read qajev.com/llms.txt: plain-text steps to check whether QAJev is installed, install it, connect the MCP server and use it.

Then paste AGENT_PROMPT.md into your agent's instructions. The agent can now crawl, check, run your project's objectives, read reports and screenshots, and follow or stop runs:

Documentation

Getting started · Writing tests · Reading a report · Projects · Jobs and top · CLI reference · MCP · Agents · Games · Mobile · Safety · Configuration · How it works · Troubleshooting

Requirements

• macOS or Linux, Python 3.12+, Google Chrome or Chromium.

• One key for Jev: TypeSafe or OpenRouter (not needed for the smoke crawl).

• For games: Godot 4, or the game's Electron app.

Limits

Jev reads pages as text and works with standard controls. Shadow DOM, iframes, canvas-drawn interfaces, file uploads, pop-up windows and custom keyboard widgets are out of its reach today, and it does not scroll far on its own (start a scenario close to its target). Pages that never stop changing make its decisions go stale; those come out as harness, never as a product failure.

On the iOS Simulator, QAJev tests native apps, but cannot yet read a web page inside Safari (its accessibility reader sees only Safari's own controls), so --real-devices ios comes out harness with that reason. Android Chrome works.

Tested on: macOS and Linux (Debian, in Docker) with Chrome or Chromium, desktop and phone view; the iOS Simulator (native apps); an Android emulator (native apps and Chrome); Godot 4 and Electron games; MCP from AI agents.

Contributing

Issues and pull requests are welcome: see CONTRIBUTING.md. Security problems: SECURITY.md.

License

MIT. Jev comes from jev-ultrafast (MIT), by TypeSafe and Browser Use.