The agent count is the only dial that matters: three for a quick look, ten when you want the site properly walked — and either way the request is a sentence, not a test suite you have to write first.
An MCP server that launches multiple browser-use agents in parallel against a URL and reports the technical problems they find. It is aimed at sites that were put together quickly and never systematically tested.
- A QA pass run by several agents at once, looking for UI bugs, broken links, accessibility issues and other technical problems
- Both live websites and local development servers as targets — `https://example.com` and `localhost:3000` are equally valid
- Control over thoroughness by agent count: 3 by default, more agents meaning a more thorough sweep
- A choice of headless or non-headless runs, with non-headless the default so you can watch
Python 3.11+, and a Google API key for the `gemini-2.0-flash` model that drives the agents. Set up with `uv venv`, `uv pip install -e .` and `playwright install chromium --with-deps --no-shell`. In Claude Code, register it with `claude mcp add vibetest /full/path/to/vibetest-use/.venv/bin/vibetest-mcp -e GOOGLE_API_KEY="your_api_key"`; in Cursor, point the command at the same `vibetest-mcp` binary and put the key in the `env` block. Built on the browser-use agent framework.
One command plus a key — uvx vibetest, then supply credentials