A four-layer local agent that turns plain-English test scenarios into executed Playwright browser actions, with no per-scenario test code. Qwen3 8B on Ollama, wired to Chrome via Playwright MCP and LibreChat, self-hosted in Docker. Teams in regulated industries cannot send test data or credentials to a third-party model provider. This stack runs fully offline at zero API cost, so nothing leaves the machine. A strict tool policy - one action per step, semantic locators over XPath or CSS, element checks before each interaction - eliminates hallucinated selectors and makes runs repeatable.