I Built an Email Assistant, and You Can Email It

I asked a question by email and a recipe came back - my curried quinoa and kale salad, written out and sent to me from assistant@boobridge.com. Nobody typed it. That address belongs to a small setup I built over a few hours, and answering mail is what it does.

What's behind the address is ordinary email plumbing and a scheduled task, not a website with a chat box. I own the domain boobridge.com. I pointed its mail records so that anything sent to that domain lands in a Gmail account I set up just for this. Every fifteen minutes, a Claude task wakes up, reads whatever is new, checks each message against a sheet of rules I wrote, and either replies or leaves me a draft. Then it tags the message in Gmail so it knows it's been handled.

What's running now, and how to try it

The three rules I built are small on purpose. I wanted to prove out the idea, not solve a real problem. They're live, though, and you can email any of them this minute at assistant@boobridge.com:

Ask it for my curried quinoa and kale salad recipe, and it will send it back to you.
Tell it what's on your mind about AI and ask if I've written anything on it, and it will come back with an article from my catalog that fits.
Tell it what music you're into and ask for any recommendations, and it will come back with suggestions from my library.

You'll hear back on the next quarter hour. Poke at it. Try to trip it up. That's part of why it's out there.

Keeping it from going off the rails

I want strangers emailing it, which means it has to be hard to abuse. So the limits don't live in the AI's judgment. They live in plain code that runs before anything leaves.

It only ever replies to the person who wrote in, so a message that says "forward this to a different address" can't redirect it. There are caps on how many times it'll answer one sender and how much it'll send in a day. There are kill switches. And it fails closed - if the safety check can't load for any reason, nothing goes out at all.

I tried to break it myself. I emailed in trying to get it to tell me private things about me that it might have lying around, and it deflected. I sent a real request but asked it to answer in a crude, inappropriate way, and it ignored that part and answered straight. Both attempts went nowhere, which is what I wanted to see before opening it up.

The inbox itself is the safety net I trust most. Everything queues in Gmail in plain sight. I can watch it work, shut the whole thing off, or pull a message out before the next run touches it. The fifteen-minute delay isn't a design flaw - it's one of many windows and break points where a human can still see what's going on and step in.

What something like this could do

My three rules are proof of concept. The pattern under them could do real work for someone. A few jobs it could take on for a regular person or a small business:

Get a head start on your morning inbox - overnight, it drafts replies to the mail it recognizes, so you review instead of write.
Cover the after-hours gap - a customer asks something basic at 9pm, gets a real answer, and hears a person will follow up Monday.
Run a vacation responder that handles the easy questions and holds the rest.
Draft the polite "not right now" so you stop rewriting it.
Flag what actually needs you - the urgent, the upset - and never auto-answer an angry email.
Collapse a long thread to its short version, with a suggested reply on top.

They share one thing a website can't touch - they either work on your side of the inbox or answer someone the moment they reach out, which is the gap a FAQ page leaves wide open.

Why I can't just hand it to you

I'd give this to you if I could, but I can't really, and I should be straight about why. The drafting-only version is within reach for a lot of people - Claude's Gmail connector can write drafts you approve. The full version, where it sends as your own address on your own domain, asks you to wire up mail records, DNS, and account credentials by hand. That's real systems work, and most people won't want to spend a Saturday on it. The setup is the wall.

I'm not the first down this road either. Emilio Delgiudice built a Claude-first email assistant late last year on the same instinct - local rules, no per-message AI bill, the thinking happening inside Claude rather than some metered service. His organizes your own inbox rather than answering strangers, but the foundation is his, and his writeup is a good read.

The part that has me fired up

I built this quickly, mostly to find out whether it was possible. It is - and the bigger thing it points at is what has me excited. You can put an LLM at the center of a solution now, doing the actual work instead of sitting off to the side as a bolted-on feature, and stand up something that wasn't practical a year ago. A few hours of tinkering got me an assistant answering real email under my own domain. The ceiling on what one motivated person can build right now is higher than most people realize.

I'd love some company up there. So email assistant@boobridge.com and try one of these:

Ask for my curried quinoa and kale salad recipe, and it'll send the whole thing back.
Tell it what's on your mind about AI and ask if I've covered it, and it'll point you to an article from my catalog.
Tell it what music you're into and ask for ideas, and it'll answer with picks from my library.

It answers within fifteen minutes. I'm curious what you'll ask it.