Artel QA Lite
Experimental. Artel QA Lite works and is improving quickly. Expect rough edges, and tell us about them in Talk to the developers.
Agentic QA for web applications. It reads your application's source, tests the running application in a real browser the way a user would, and writes a report a QA lead can act on. It runs on your machine, on your own Claude Code subscription.
How it works
- Identify. It reads your source and lists every control a user could act on — each field, button, link and label — with the page it is on. That list is what coverage is measured against, so it comes from your code, not from whatever a crawler happened to find.
- Walk, then inspect. It opens every page, signs in, and records what is really there. Then agents, working in a real browser, use each control as a user would.
- Judge by evidence. An agent never writes its own verdict. Artel measures what actually happened — the screen before and after, every request that left the browser — and the verdict comes from that. If an agent says "saved" and nothing was sent, the report says so.
Your first session
From the root of the application's repository. Everything before the last two steps is free.
metantel doctor # check this machine has what it needs
metantel init # then set your app's address (and sign-in) in .metantel/config.json
metantel identify # what your source declares
metantel plan --open # the test plan, as a page
metantel sweep # walk the application: what is really on each page
metantel run --budget 5 # the first real run, capped at five dollars of your subscription
metantel report --open # the reportThe console
metantel artel-lite opens a local console in your browser. It shows the stages of the current run as they happen, findings as they are filed, and coverage. From it you can start a run, set priorities, or stop a run.
The report
Every run writes a report as a web page (it prints well) and as Markdown (paste it into a ticket). At a glance:
| Controls | Every field, button, link and label your source declares |
|---|---|
| Reached | Opened by an agent or by the walk |
| Assessed | A pass or a fail, with evidence |
| Passed | Behaved as its label says |
| Findings | Defects, with severity, each with what was seen |
Then the findings, results per control, what could not be reached and why, and the test cases.
Cost
Agents run on your own Claude Code subscription; Metantel does not bill for a run. Reading the source, the plan, the walk and a dry run cost nothing. On real applications a run costs about $0.04–0.07 per control assessed with the default model.
--budget USDstops the run at that spend; the console shows spend as it happens.--target Pstops once that share of controls is assessed.
What it needs
- The Metantel CLI (install), with Node.js 20 or newer
- Claude Code, installed and signed in
- Google Chrome or Chromium
- Your application's source, and the application running somewhere it can reach