Kiwame
An AI QA engineer that runs inside Claude Code, Codex or Cursor. Your coding agent plans the tests and drives a real browser; Kiwame supplies the workspace, the evidence and the safety gates.
- Status
- Released, installable as a plugin
- Built with
- Python, Playwright, Claude Code, Codex, Cursor
- Links
- Source
Kiwame (極め) is the Japanese word for an expert’s final verdict: the certificate a master appraiser wrote to prove a sword genuine. The tool gives a verdict per test step, with video and screenshots to back it up.
The idea
Most AI testing tools are a product with a model bolted on, and you pay for the model twice. Kiwame never calls a model itself. Your coding agent does the thinking, so the whole thing runs on the Claude Code subscription you already have.
What Kiwame provides is everything around the model: a workspace, a real browser, captured evidence, and the guard rails that make it safe to let an agent loose on a live application.
You stay in charge
- Nothing runs unapproved. You approve every scenario, and the approval is bound to the file. Edit the file and it needs approving again.
- Anything irreversible asks first. Delete, pay, send: that’s a native OS dialog on your screen, which the agent cannot answer for you.
- Passwords are never seen by the agent. It types a placeholder; Kiwame substitutes the real value inside the browser.
Install
/plugin marketplace add sarangkkl/nkqa
/plugin install nkqa@nkqaThe command is kiwame, with qa as a short alias.