Enable both snapshot and screen capture tools

[clharman]

Both of these are extremely useful in certain circumstances - the snapshot for navigation, and the screenshot for visual analysis. Currently activating the --vision flag enables the screen_capture but disables the snapshots.

[pavelfeldman]

You can capture screenshots in snapshot mode.

[wernerthiago]

Nowadays, if you use the config file it is possible to make a server connect to other, so we can have both versions. However, it does not work well. It looks like vision is looking for a tab and does not recognize tabs opened with default server. So vision keeps trying to open about:blank tabs to fill-out the expectation 😅.

If we simply launch a vision mode, things that were previously working such as log in etc stop working.

I believe and I hope that this integration is going to be improved in the future.

[clharman]

@pavelfeldman seems like the non-vision mode screenshot does not return the image in the body, it just saves it.

Example:

Non-vision:

(The screenshot body returned just includes the file location, and then the page snapshot - it does not include an actual image)

Vision:

(It actually returns the image in the body)

This means whenever I want the client using the MCP to be able to look at the image, I have to switch to vision model. But then, when I want the snapshot, I have to switch back. Or am I understanding it improperly?

[pavelfeldman]

Which Client is that? Playwright will not include image data for Cursor since it does not support image mime type, but the rest of the clients should be getting it. It could be that yours drops it on the floor.

[clharman]

Yep it's Cursor - however the vision mode browser_screen_capture seems to work great with Cursor! (I was also using v0.0.20 previously and it worked well then, too)

[pavelfeldman]

I see what you are saying - when I select sonnet model it does interpret the image just fine. The problem is that we can't tell which model Cursor uses and by default it often picks gpt-4.1

[clharman]

Gotcha, makes sense. Could a workaround be to keep enabled the snapshot tool when --vision is active? That way the screen_capture tool from the vision package could supply the images, but the snapshot from the non-vision package could still supply the snapshot (which is, as I'm finding, more helpful in 80% of cases, as your docs state).