Why we built Krewva — and why every chatbot stops short

AI assistants that close their laptop at 5pm don’t help anyone.

That’s the sentence I wrote down at 2am, three weeks before we started building Krewva. I’d been awake since the previous morning, working through a backlog of customer emails for a project that had nothing to do with AI, and I’d done what I’d been doing for two years at that point: I had ChatGPT open in one tab, my email open in another, and I was copy-pasting between them. Draft a reply. Paste into the chatbot. Adjust the tone. Paste it back. Send. Move to the next email. Repeat thirty-eight times.

Around message twenty I stopped to think about what I was actually doing. The chatbot had not, in any meaningful sense, replied to my emails. I had replied to my emails. The chatbot had been a slightly faster typewriter. It had no memory of what I’d said to this customer last month, no context for the project, no idea whether the person on the other end was a hot prospect or a chronic complainer. It was, fundamentally, a more eloquent autocomplete with a fancier UI.

And the moment I closed my laptop, it stopped. No emails got drafted overnight. No threads got triaged. No replies went out. I’d built a tiny, expensive, 9-to-5 employee that lived in another tab.

That’s the problem with chat-window AI. It’s an assistant in a tab.

What chat windows are actually for

Chat windows are a great UI for one thing: a single human asking a single question and getting a single answer. ChatGPT, Claude, Notion AI, Cursor — all of them are optimized for that interaction shape. Type a prompt, get a response, copy what you need, move on.

That’s a fine pattern for “help me debug this regex.” It’s the wrong pattern for “handle the inbox the way I would handle the inbox.” Inbox work is fundamentally not single-shot. It’s continuous, multi-thread, multi-context, multi-relationship. Your dad sending you a meme requires a totally different reply than your VP sending you a P0 incident report. And the work doesn’t stop when you close the tab. The emails keep arriving. The Slack threads keep growing. The iMessage backlog keeps stacking up.

So what we got from the chat-window era was a generation of AI products that all said the same thing: “we’ll save you time on emails.” And what they all delivered was the same workflow shift: open another tab, paste your context in, paste the answer out. Pure context-switching tax dressed up as productivity.

The shape we kept landing on

When we started talking about Krewva — Haiyang and I, late nights, whiteboards, margin notes — we kept landing on a different shape.

The shape was: don’t ask the user to come to the AI. Send the AI to where the work is.

If the work is in Gmail, the agent should read Gmail, draft replies in Gmail, and surface a feed of decisions to approve, not paragraphs to copy. If the work is in iMessage, the agent should read iMessage on the user’s own laptop (because iMessage doesn’t have a real API and the chat database lives locally), draft a reply, and queue it for the user’s review. If the work is in Slack — same thing. WhatsApp — same thing.

This is what we mean by inbox-native. It’s a constraint, not a marketing phrase. Krewva does not ask you to forward your emails to a magic address. Krewva does not run inside a chat window with a prompt box. Krewva sits inside the apps you already use, on the surfaces you already trust, and reads the same threads you’d be reading.

The cost of this constraint is real. Inbox-native means we have to care about the things you’d hope your AI assistant doesn’t have to care about: Apple’s Full Disk Access flow, Google’s OAuth verification, WhatsApp Web’s quarterly DOM changes, Slack’s bot scopes, Cognito session refresh, push notification deep-links. Every connector is a separate engineering project. Every platform is a separate trust conversation with the user. None of that ships in a “wrap an LLM” weekend.

But the upside is the entire reason this product exists. When you wake up in the morning, your messages have already been triaged. The replies you would have written are already drafted. The threads that needed context have already had that context surfaced. Instead of opening Gmail and seeing 47 unread, you open Krewva and see 47 decisions — most of them already drafted, ready for you to swipe approve.

The 24/7 thesis

Here’s the thesis behind all of this, written as plainly as I can: humans don’t operate 24/7, but our messages arrive 24/7.

Your customers are in different time zones. Your friends text you at midnight. Your kids’ school sends emails at 7am. Your boss pings on a Saturday. The inbox doesn’t sleep. So the assistant that handles the inbox shouldn’t sleep either.

Every chatbot — every chat-window AI — sleeps. The moment you close the tab, nothing happens. Krewva is the opposite. Krewva runs in a worker process on a server we own, reads your connectors on its own schedule, and works through the queue while you’re at dinner, while you’re asleep, while you’re on a flight with your phone on airplane mode. When you open the app the next morning, the work has been done.

That’s what “AI employees who work overnight while you sleep” means. It is not a marketing slogan. It’s an architectural choice. The architecture is: the agent runs on a server, not in a tab.

If we’d built Krewva as a chat-window AI, that thesis is impossible. The user has to be present for the work to happen. With Krewva built as inbox-native, on a worker, with persistent connectors — the work happens whether or not you’re around. Your job becomes approval, not authorship.

Why I personally needed this

I’ll be honest about the personal angle, because I think it matters.

I burned out on email. Not in a dramatic way, just in the slow, grinding, accumulated way that founders burn out on things. By the time we started Krewva I was opening my inbox in the morning with low-grade dread. I knew there were forty new threads. I knew most of them were variations of three or four emails I’d written before. I knew that if I sat down for an hour I could clear them — but I also knew that doing that, every morning, for the next ten years, was a tax on my life that I didn’t want to pay.

The honest version of “why I built Krewva” is that I wanted my mornings back. I wanted to wake up to a feed of approve / approve / approve / send this one with a tweak / archive this — instead of a wall of unread. I wanted the work that had to happen to actually happen, but I wanted to stop being the bottleneck for every keystroke of every reply.

A chatbot does not solve that. A chatbot is more keystrokes — your prompt to the chatbot, then your tweak of the chatbot’s reply. An inbox-native crew that drafts in your own voice, queues for your approval, and earns autonomy contact-by-contact — that’s the shape of the actual fix.

What we mean by Krewva Life and Krewva Biz

One last thing about the name. Krewva is a play on crew, but the product split matters.

Krewva Life is one unified agent for your personal apps: messages, email, calendar, Drive, family coordination, money and portal work, all flowing through one approval deck. You should not have to think about which internal specialist owns a personal task. You should see the decision, approve it, and move on.

Krewva Biz is where the specialists show up: eight standalone department agents for sales, marketing, finance, legal, HR, ops, PM, and analytics. They work through the same approval-card primitive, but the interface is a business office: walk in, click the department lead, review the pending stack, approve, deny, or redo.

This matters because “AI assistant” is the wrong noun for both products. Krewva Life compresses your decisions into one daily flow. Krewva Biz compresses business labor into eight always-on functions. The bet is that by the time most people meet Krewva, it doesn’t feel like an AI tool at all. It feels like work already moved forward.

Closing

I started this post saying that AI that closes its laptop at 5pm doesn’t help anyone. I’ll close by saying: AI that lives in a tab is the same thing in a different shape. The tab closes. The tab closes when you sleep. The tab closes when you’re on the train. The tab closes the moment you go do anything that isn’t sitting at a computer.

If you build the assistant inside the tab, you’re capping the entire product at the user’s screen time.

If you build it inside the inbox, on a server, on a 24/7 worker — the cap is gone.

That’s why we built Krewva. That’s the whole pitch. Everything else — the trust dial, the voice profile, the connector engineering, the bucketing — is just the work it took to make that pitch survive contact with reality.

We’ll write more about each of those pieces in the coming weeks. If you want to be on the list for the early app, the link is in the footer.

— Zeming