How about using playwright-mcp as a tool for WebSurfer agent?

[ksachdeva]

https://github.com/microsoft/playwright-mcp

Hi,

Am wondering if there are any limitations in playwright-mcp that would come in the way of WebSurfer implementation?

Regards & thanks
Kapil

[mmurad2]

@ksachdeva The existing Playwright-controller implementation was explicitly tuned on the expected behaviors of the controller. It supports some neat UX features like taking over control, seeing the mouse hovering action, etc which likely don't come out of the box for playwright-mcp.

cc: @afourney

[afourney]

That's correct. Our controller predates the playwright-mcp by about a year, and it is pretty deeply integrated at this time. We use set-of-mark prompting + DOM + accessibility tree to ground actions. There is much overlap with markitdown-mcp, but it's not complete. For example, set-of-mark prompting allows many multimodal models to work with screenshots -- not just the computer-use models.

I would be happy to investigate this more in the future, but it's a little more complex than just swapping controllers.

[ksachdeva]

thanks @afourney @mmurad2