Most ways of letting an LLM drive a browser hand it a screenshot and ask it to guess where to click. That burns vision tokens and breaks the moment a layout shifts. This server takes the opposite route: it feeds the model Playwright's accessibility tree, so the page arrives as named, structured elements the LLM can reference directly. Actions become deterministic rather than coordinate-guessing — which is also why the same logic powers self-healing tests.
What Sets It Apart
- Accessibility-tree snapshots instead of pixels mean no vision model is required and token cost stays low — the model works from text the way a screen reader would.
- Clicks and types target semantic elements by role and name, so they survive layout changes that would defeat coordinate-based automation.
- It runs as a standard MCP server, so it drops into VS Code, Cursor, Claude Desktop, and other MCP clients without per-client glue code.
- Persistent browser context lets one session carry state across many steps, suiting long-running autonomous workflows rather than one-shot scrapes.
Who It's For And The Trade-offs
A great fit if you want an agent that explores web apps, repairs flaky tests, or runs multi-step browsing reliably across a stable, accessible DOM. Look elsewhere if your target relies on canvas-rendered or visually-encoded content with no accessibility metadata — there the structured-snapshot approach has little to read, and a screenshot-based tool may capture more. It also assumes you are comfortable running a browser process alongside your agent.
