Most browser servers give you one of the two: either a scripted runner or an interactive session. Here the same platform does both, and the dev session runs single-step by default, so the model proposes one action and you see the result before the next. Worth knowing before you plan a long run: a session lives 30 minutes, ten can be open at once, and one MCP call has 25 seconds to return.
The MCP mode of qa-use, a browser-testing tool built on the desplega.ai platform. Tests are YAML files whose steps read like a checklist — goto, fill, click, to_be_visible — and the same engine will also hold a live browser open so you can steer it a step at a time.
- start_dev_session opens a browser on a URL and keeps it alive for the rest of the conversation
- interact_with_session with action "respond" takes an instruction in ordinary English — "click the submit button", "fill the email field with test@example.com" — and runs exactly one step per message
- The same call accepts "close" when you are done with the session
- monitor_session reports what happened: the current page URL, the task history, and whether the session is waiting on you
- Test steps cover navigation, input, waiting and assertions, plus ai_action, ai_assertion and extract_structured_data for the parts a fixed selector cannot express
The npm package @desplega.ai/qa-use, version 2.19.0, started as `qa-use mcp` for stdio or `qa-use mcp --http`. It needs a desplega.ai API key in QA_USE_API_KEY and an app config id for the site under test; QA_USE_FORCE_HEADLESS pins every launch to headless.
One command plus a key — npm install -g @desplega.ai/qa-use, then supply credentials