[berkes]

> Claude's webpage on a side pane, with all the issues of a squished webpage.

I have used that feature for a few weeks now and find it utterly useless.

Partly because it is squished. But mostly because it offers no value over just having a tab open with Claude (or in my case Mistral).

The extra buttons (summarize) and integration (context menu) hardly ever work (pages and selections are often too large for gpt, copilot, mistral or even claude and the sidebar just gives an error) but even if they did: what problem do these extra buttons and integrations solve? Am I missing something?

Do note that I would love integration the other way around: to have an AI agent (through an MCP for example) drive my firefox. Safely, contained, etc etc. I am not an AI luddite. I just find the firefox sidebar offering no value at all.

[chrisweekly]

I think that's exactly what people are doing w the Playwright MCP...

[berkes]

It's indeed exactly what I do.

And I think Firefox should step up and become a better alternative to playwright. One geared at developers with tools like profiling, dom-manipulation (basically the Developer Console) assertions, visual comparison etc. Or (and?) one geared towards normal browsing and interacting with webapps.

[prophesi]

I was also frustrated by the short context length for summarizing and found you can increase it in about:config via:

browser.ml.chat.maxLength

[berkes]

I googled* it and dod the same. But that merely moves the problem forward. Now the API returns errors that it's too large.

A smarter client would use a tool chain in which the first step is a model that's good at taking large contexts/data and extracting actual content from it. Many sites have a very low S/N ratio (readable content / dom).

Then pass that content, eg markdown, along to a model that's optimized at getting relevant parts out of content for the task at hand.

And only then onto the generic model to "do stuff" with it.

But many clients, including afaiks the Firefox one, just send the entire dom or html along to a generic model.

*Well DuckDuckGo

[prophesi]

Sorry for the late response; from what I can tell from my experience is that they're good at eliminating the unnecessary information (at least, as best as it can). You can see exactly what is passed in to the chat bot (for me, openweb-ui) and can see it's all text elements of what makes pertinent sense to be summarized.

My ideal would be an additional prompt that says "hey this page content is beyond your context length, fine with extending it to whatever the maximum context length plus some for the model's context length"?

[trenchpilgrim]

Being able to e.g. hook up Claude Code to any webpage would be killer for both web development and task automation.

[berkes]

I already do this via docker desktop MCPs in which I run a.o. Playwright¹.

Zed and VS code can now drive a browser. E.g. to perform tasks like "run the site on a local server and verify that you see the counter increase when you click this new plus-button"

¹https://hub.docker.com/mcp/server/playwright/overview

[trenchpilgrim]

Thanks, I'll try this in Zed!