
If you’ve used Playwright MCP for more than just demo logins from YouTube, you’ve probably run into this issue: the agent misses some elements on the page, gets confused, or completely loses context.
The reason - Playwright MCP sends an ARIA snapshot to the LLM, not the full list of interactable elements from the DOM.
Together with my team, we built an MCP upgrade that:
- serializes the full DOM tree
- returns all interactable elements
- provides a complete page context
As a result, the agent gets a full picture of the page, understands how to interact with elements, and can generate significantly more accurate and comprehensive test scenarios from the first attempt.
https://github.com/MobiDev-Org/treegress-browser-mcp (open source)
Hope you find it helpful. I’d really appreciate your feedback