The comparison the project draws with a Playwright-based server is that the agent never has to locate the element first — it states the action and the control algorithm resolves it, saving both the round trips and the tokens spent describing the DOM. The same property means your steps survive a redesign that would have broken a selector.
A browser automation server built on kogiQA's browser control algorithm: the agent says what to do in natural language and the browser does it, without CSS selectors and without a vision model reading screenshots.
- Pages driven by description — log in, add a product to the cart, check a badge — with no selector to write or maintain
- Console errors collected across the pages you visit, so the agent can go and fix them in the source
- Layout and style problems inspected at a viewport width you name
- A site explored and documented: follow internal links to a depth and produce a Markdown description of each page and its features
- End-to-end tests written from a walkthrough of the real flow, covering the happy path and the validation errors
- Exploratory testing — edge-case inputs into every form it finds, with a report of what looked broken
Node.js 20 or newer and an MCP client. Auto-install with `npx kogiqa-mcp@latest`; Claude Code takes `claude mcp add kogiqa-browser npx kogiqa-mcp@latest`, and VS Code, Cursor and Windsurf all accept the same command.
One command — npx kogiqa-mcp@latest