Let Your AI Work Behind Your Logins:Log In Once, Then Draw One Hard Line
Most AI assistants stop at the login screen. Log in once yourself, let the AI work inside that session, and make it stop before anything it can't undo.
Most AI assistants hit a wall the moment a job needs you to be logged in. The fix I landed on is simple: log in yourself, once, let the AI work inside that session, and make it stop before anything that spends money, signs your name or sends a message.
I got there by building my own browser tool on a Sunday afternoon. Here's the story, what the tool does, and the rule I'd hand to any business owner.
The wall I hit in September
I use an AI assistant to find Reddit threads where I can help and to draft my replies. Every reply waits for my yes on the exact words.
In September, Claude's Chrome extension started refusing every Reddit task, citing its terms. My Reddit account was fine. The refusal came from the assistant's side, before anything reached Reddit.
That was annoying for one site. But it pointed at a bigger problem. A lot of my work, and probably yours, lives behind a login. If the AI can't get through the front door, it can't help with anything inside.
A crazy idea: skip the cloud
Around then I heard about Browserbase, a service that runs browsers in the cloud for AI agents. Your agent connects to one of their browsers and does its browsing there.
Then I had a crazy idea. What if I didn't need the cloud at all, just my own browser that my AI could drive? My own Mac, my own logins, nothing sitting on someone else's server.
So on Sunday, October 25, I had my agents research open-source tools that already did most of the job. The best fit was agent-browser, a free tool from Vercel Labs with about 44,000 stars on GitHub. It turned out it was already installed on my Mac. It did most of what I needed, so I started there, added what I was missing, and called the result bb-local.
By that evening, Claude Code could use it.
What it does now
Three things matter, and none of them are clever. They're the boring parts that make it safe to let an AI into your accounts.
Log in once. I sign in to a site by hand, in a window it opens for me. It saves that session, encrypted, on my own Mac. After that, every AI assistant I run starts already signed in. I never paste a password into a chat, and I don't log in again for each tool.
Hand off the hard parts. When a site asks for a code from my phone, it stops and waits for me. If I'm away from my desk, it sends me a link, and I finish that step from my phone. The AI never pretends to be me at the moment a site is checking that it's me.
Check its own work. When it fills out a form, it fills the plain fields, then reads each one back from the page to confirm it stuck. That catches the quiet failures, like a box that reformats a phone number or a dropdown that didn't take. It will not touch passwords, signature boxes or the Submit button. Those wait for me.
The first real job wasn't Reddit at all
Three days later it did a job I hadn't planned for. My running team's site was an old WordPress install that needed to become something simpler to keep up.
The usual way is to export everything from WordPress, hand the files over, and hope nothing gets lost. Instead, I logged in once, and an agent pulled every page and rebuilt the site as a plain static site. I never exported a single file.
The tool was built for one annoying site. The value showed up on a completely different one, because the real unlock wasn't Reddit. It was letting the AI work inside a login I control.
The rule for business owners
Most of your business sits behind logins too: your scheduling tool, your supplier portal, your website. Your AI can only help with what it can open.
You don't need to paste a password into an AI chat to open those doors. You need a setup with three properties:
- You sign in yourself, once. The AI works inside your session. It never holds or types your password in a chat.
- Anything that proves it's you comes back to you. Codes from your phone and security checks are your job, not the AI's.
- One hard line. The AI stops before anything that spends money, signs your name or sends a message. It can read, fill and check. You press the button.
That third rule is the one I'd start with even if you never build anything. Write down, today, the three actions in your business that can't be undone. Those are the ones your AI should always bring back to you.
Which site do you log in to every week that you wish someone else could handle?
When an AI Says the Job Is Finished: 4 Checks I Use After Testing Gemini 4 Argon Against Claude Opus 5.5
If an AI told you a job was finished and it wasn't, a smarter model won't fix it. Here are 4 checks I use after testing Gemini 4 Argon against Opus 5.5.
ReadHow Claude Code Set Up 16 AI Assistants for My Business in 18 Minutes, and the 4 Rules From the 5 Days of Planning Before It
Claude Code set up 16 AI assistants for my business in 18 minutes. The 5 days of planning before it gave me 4 rules you can use for any new helper.
ReadHow I Connect My Macs With Tailscale So My AI Agent Can Work on an Always-On Computer
My AI agent works on a Mac that stays on all day. How I connect my machines with Tailscale, Screen Sharing and Remote Login, and how you can too.
ReadWorking on something like this?
Bring the app or the process to a free 15-minute call. I will tell you what I would look at first, and whether I am the right person for it.
