This is preview documentation for the upcoming release. Publication is pending live Jev validation. The demo uses a live navigator with mocked Jev element selection.
WebTest AI
Write a browser test as a Markdown goal and explicit success checks. A navigator chooses actions from fresh observations; the runtime executes bounded actions and evaluates the checks. Reports show the verdict, browser evidence, model identity, available token usage, and cost estimates.
Start here
- Install the beta in a clean project.
- Run Everyday Goods, then connect a live navigator.
- Write your first test for your own site.
- Read the report and quality checks.
The default demo is a scripted rehearsal. Use session mode for an external agent, or configure a hosted or local model for a live run. The demo mocks Jev element selection. See model configuration.
How it works
The navigator decides what to do next. Jev, when configured, resolves a target to an observed element; it does not plan the test or decide the verdict. Success checks stay fixed throughout the run. Missing usage or pricing is reported as unavailable rather than zero.
Accessibility findings, lab LCP/CLS, and configured visual checkpoints appear separately from functional results. A functional pass can still include quality warnings.
Go further
- Commands and configuration
- Long journeys, frames, tabs, and saved login
- Agent workflows
- Migration from the previous beta
- Changelog
This documentation targets the 0.2.0-beta.1 candidate. Before publication, install the supplied tarball. Release status and installation instructions are in Try the beta.