My unread count hit 3,847 on a Tuesday in March. I remember because that’s the day I screenshot it to send to a friend who thought I was exaggerating about my inbox.

She wasn’t wrong to doubt me. Everyone says they’re drowning in email. And most people mean they get 60 real messages a day and don’t love the work of reading them. But mine had become something else: a graveyard of half-started threads, calendar ghosts, payment reminders from SaaS tools I’d stopped using, and somewhere in the middle, a client contract I was supposed to sign last week.

I’d tried the obvious things. Superhuman. Shortwave. Boomerang. A period in 2024 where I paid someone on Upwork to triage my inbox for two hours a day. She quit. And she also thought I was exaggerating. None of it stuck because none of it addressed the actual problem, which was that I hated the act of looking at email, and every app that promised to help still required me to look.

The OAuth wall

Martin Fowler wrote a piece in February called “Agentic Email” that described exactly what I wanted: an AI that would read every message, decide which to ignore, and draft replies for the ones that mattered. He also said it was terrifying. Granting an LLM full OAuth access to the nerve center of your digital life is the kind of thing that sounds clever in a demo and catastrophic on the day a password-reset email gets intercepted by a prompt injection.

So I didn’t do it. For four months I sat on the problem because every “AI email assistant” I could find wanted a Gmail OAuth token, a forwarding rule, or my actual credentials piped into a third-party server I’d never heard of.

Then I installed dassi

It doesn’t connect to Gmail. No tokens. No forwarding rules. It looks at what’s on my screen, same as I do, in the tab I already have open.

Six days of drafts

Day one I opened Gmail with the dassi panel on the side and asked it to draft a reply to a vendor who had been chasing me about a renewal for three weeks. It read the thread, noticed I’d already declined verbally in a previous email, and drafted a polite “I thought this was resolved on March 14. Can we confirm the cancellation went through?” in about eleven seconds. I clicked send. That email had been open in my head for a month.

Day two I batched 40 messages. Scroll to the top of a thread, hit the side panel, say “draft a short reply confirming the meeting and asking for the agenda,” get something sensible. Sometimes I tweaked it. And sometimes I didn’t. The model behind it was GPT-5.2 running on my own ChatGPT subscription, no API key, no separate billing, nothing extra I’d have to explain to accounting.

By day three the volume started dropping because I was actually finishing conversations instead of letting them rot. Day four I wrote three-word prompts. “Polite no.” “Ask for date.” “Say I’m traveling.” The agent would read the thread and produce something that sounded like me, because it was working from the exact context I was.

Day six, zero unread.

I don’t say this to brag. I say it because six days is what it took after four months of paralysis, and the thing that broke the logjam was not a new app or a new habit. It was the part where I stopped needing to log a third-party service into my most sensitive account. Other people have written about the same shift for Gmail without OAuth, and the pattern is identical: the assistant sees what you see because it lives where you work.

What it doesn’t do

It doesn’t send. It drafts, I read, I hit send. Some people go further and let the agent autonomously reply to low-stakes messages, and I understand the appeal, but Fowler’s security argument lands hard for me. Reading is cheap. Sending from my account is the part with consequences, and I want my eyeballs on it.

It also doesn’t handle emotional nuance worth a damn. A sympathy note to a colleague whose dog died, I wrote that myself. An email firing a contractor, also me. But the 200 other messages in a week are not like that. The “yes I can meet Thursday” and “forwarding your invoice to finance” and “thanks, this was really helpful” ones. Those are not writing. Those are typing. And I don’t want to spend my Tuesday typing.

The thing nobody tells you about inbox zero

The inbox-zero people will tell you the fix is discipline. It isn’t. The fix is fewer keystrokes per decision, which is boring to say but turns out to be the whole game. I’d been treating email like a skill problem for a decade when it was a bandwidth problem, which is why every productivity app in my graveyard failed. They made the typing prettier and kept the typing.

If you want to try it, dassi is on the Chrome Web Store. It’s free. It doesn’t store your emails. And unlike every other inbox tool I’ve paid for, it’ll still work after the next Gmail redesign, because the only thing it depends on is the tab I already have open.

I still get 3,000 messages a week. I just don’t read most of them now. Which was apparently the goal the whole time.