Let an AI agent test your web app as multiple users (MCP)
Short answerQA SessionLab ships a zero-dependency MCP server. Register it in your AI client with the token from Settings → AI control, and the assistant gets 27 tools to create isolated sessions, open pages as a role, read/click/type, compare roles and record bug reports in your real Chrome.
Step by step
- Copy the tokenOpen the QA SessionLab side panel → Settings → AI control and copy the token.
- Register the MCP serverclaude mcp add qa-sessionlab -e QSL_TOKEN=<token> -- node /path/to/qa-sessionlab/tools/mcp/qsl-mcp.mjs (other clients: run the same command with QSL_TOKEN).
- Ask for a role test“Create Admin and Customer, log in to http://localhost:3000 as each, compare /orders and report permission differences.”
- Review the evidenceThe assistant returns what each role saw, screenshots and Markdown bug reports with console errors and failed requests.
Why this works
- Generic browser automation gives an AI one identity; role testing needs several isolated logins at once.
- qsl_compare_roles returns structured, comparable results instead of screenshots to interpret.
- The bridge listens on 127.0.0.1 only and uses mutual HMAC authentication; the token never crosses the socket.
Questions
Which AI clients work?
Any MCP client: Claude Code, Claude Desktop, Cursor and others that can start a stdio MCP server.
Does the AI see my normal browsing?
Only what you ask it to open through the tools; your normal profile is not part of any session.