connecting

Settings

Keys are held in this process's memory only, and forgotten when it closes — unless you tick “remember in my keychain”, which puts them in your operating system's keychain so they come back after a restart or an update. Nothing is written to the project folder, so a new build cannot take them with it, and nothing leaves your machine except calls to the provider you configure.

Model providers

You need at least one. Without any, the agent falls back to a deterministic offline stub that cannot write usable test steps.

Anything that speaks the OpenAI API — Groq, Together, DeepSeek, Fireworks, Mistral, or your own vLLM or Ollama. Its models appear in Which model does what alongside the rest.

Which model does what

Seven jobs, each with its own model. Actor runs on every browser step and drives most of the cost; reasoning makes the judgement calls and deserves your best model. Saved to — configuration, never keys.

When to write cases

Before writing anything, the agent scores how testable a ticket is. Below this score the run stops and reports the requirement gaps instead — the questions for the Product Owner are the output. Lower it to get cases from a rough ticket; the gaps are still reported either way.

Where the test-case sheet goes

After a ticket is analysed, the sheet has to reach whoever asked for it.

Integrations

Optional. ClickUp lets it read tickets directly instead of from a file.