Start the server
node cli.cjs mcp
Speaks MCP over stdio. Whatever you are using connects to it the way it connects to any other MCP server.
What your assistant gets
- Start testing: any of the eight checks, scoped how you like.
- Read the findings, with their evidence and confidence.
- Fetch a fix prompt for a finding, so it never has to be pasted by hand.
- Run the gate, and get back whether anything new is ship-blocking.
- Verify a fix in a browser, which is the part that stops the assistant marking its own homework.
Why this is the good way to use it
The loop that works is: the assistant changes something, a browser checks whether it landed, and the assistant reads the result. Without the browser in that loop, an assistant’s report that a fix worked is a claim with nothing behind it, which is the exact failure this product was built to catch in your product.
Per editor
Claude
Through the connector in the browser, or in Claude Code over stdio. Ask for a test; the fix prompts come back.
Codex
As a connector, and from the Codex CLI. Hand it a task and a browser proves the fix landed.
Cursor
Added under Settings, then mentioned in the composer. Findings, evidence and the gate arrive as tools the editor calls.
The MCP server runs the same modules with the same rules. Two of the eight need your model key; three do not, and your assistant can run those without spending anything.