Last week I needed a client.json file from Google Cloud Console. OAuth credentials for a Gmail integration. I do not use GCP daily, and every time I go back to that console I feel like I’m navigating a government building where someone rearranged all the floors since my last visit.

So I asked Dassi to do it instead.

The 23-click problem

I counted once. Setting up OAuth credentials in Google Cloud Console from scratch involves roughly 23 distinct clicks across 7 different screens, assuming you already have a project created and you don’t get lost along the way (which you will, because the “APIs & Services” section has subsections that look identical but do completely different things). You need to enable the Gmail API, configure an OAuth consent screen with scopes that sound like they were named by a committee, create the credential itself, and then download the JSON file. Every step has its own page, its own save button, its own confirmation dialog.

And GCP is not even the worst offender. AWS IAM makes you want to close your laptop.

What actually happened

I had Dassi open in the Chrome side panel. I navigated to console.cloud.google.com, told Dassi “I need to create OAuth 2.0 credentials for Gmail API access and download the client.json file,” and watched it work. It found the right project. Clicked into APIs & Services. Enabled the Gmail API. Went through the consent screen configuration. Created the OAuth client ID. Downloaded the file.

The whole thing took about four minutes. But the real win was not the time saved. It was that I did not have to figure out which of six similarly-named menu items was the right one, or remember whether “credentials” lives under “APIs & Services” or somewhere else entirely.

Every admin panel is the same nightmare

Skyvern 2.0 launched last month with some impressive browser automation demos, and there’s been a wave of open-source browser agent projects hitting GitHub. The browser agent space is having a moment. But most of the demos show agents doing things like booking flights or filling out simple forms.

Nobody talks about the admin panels.

Stripe’s webhook configuration page. Cloudflare’s DNS settings buried three levels deep. Azure’s portal, which somehow manages to be both overwhelming and incomplete at the same time. These are the web UIs that actual developers and ops people spend real hours in, clicking through screens they visit just rarely enough to forget where everything is.

And these UIs are all web pages. They run in Chrome. Which means a browser agent can operate them.

Why chatbots cannot solve this

I’ve used ChatGPT to ask “how do I set up OAuth in Google Cloud Console” and gotten perfectly accurate step-by-step instructions. Great. But then I still had to do all 23 clicks myself, alt-tabbing between the instructions and the console, trying to match what ChatGPT described to what I was actually seeing on screen because Google redesigned something since the training data cutoff.

The gap between “explain the steps” and “do the steps” is enormous. ChatGPT, Claude on its own, any chatbot that lives in a separate tab cannot click buttons on a page it cannot see. That’s the fundamental limitation of AI tools that exist outside your browser.

Dassi sits in the Chrome side panel with full access to the DOM. It sees the same page you see. So when I say “create OAuth credentials,” it can actually find the button, click it, fill in the form fields, and navigate the multi-step wizard. It is the difference between reading assembly instructions and having someone build the damn shelf for you.

The pattern scales

Once you realize an AI agent can navigate Google Cloud Console, you start seeing the pattern everywhere. I’ve since used Dassi to rotate API keys in Stripe, update DNS records in Cloudflare, and configure webhook endpoints in a SaaS tool whose admin panel looked like it was designed by someone who really loves nested accordions. Each of these would have been a 10-15 minute exercise in clicking around an unfamiliar UI while squinting at documentation.

The tasks people still do manually in their browser are not just the obvious ones like email and form filling. It is the complex admin work that eats an afternoon because you only do it once every few months and the UI changed since last time.

Browser agents that can handle these UIs are not a nice-to-have. For anyone who manages infrastructure, configures third-party services, or maintains integrations across multiple platforms, this is where AI stops being a chatbot curiosity and starts being a tool you actually rely on.