I used NanoClaw as my executive assistant for a month

A few weeks ago, I got a calendar invite for date night. Restaurant already in there, address and everything. I figured my wife had taken care of it.

She hadn't. NanoClaw had.

We went. Great spot, one we'd somehow never discovered despite living five minutes away. But that moment, realising an AI had made the booking and my wife hadn't batted an eye, is the most accurate preview I can give you of what this tool actually feels like to use. It starts acting inside your real life, and the handoff is seamless enough that nobody notices.

I ran NanoClaw for about a month: its own Google Workspace account, shared calendars, added to email threads, a structured system for tracking everything it was supposed to own. The setup alone tells you something about what it actually takes to run an AI at this level. Here's what I learned.

What worked and what didn't

Follow-ups were where it pulled its weight. At any point I had a handful of conversations sitting in a half-finished state: someone I needed to nudge on a decision, a call that needed rescheduling, a document I'd been waiting on for too long. That stuff tends to pile up quietly until it becomes a problem. NanoClaw tracked it and acted on it without me having to hold any of it in my head. The stronger move, though, was when it picked up something I'd let slip completely. Executing what you remember is useful. Catching what you forgot is the part that actually changes things.

It also caught calendar conflicts before they landed on me. Moved things around, looped in the right people, handled the coordination. That's the kind of task that isn't hard, it just requires attention at the exact wrong moment. Having that covered made a real difference.

Where it fell short was subtler. The real problem isn't that you forget to update it — it's that the system has no reliable sense of what should persist and what shouldn't. A standing priority, an ongoing relationship, a decision you made three weeks ago: all of that lives in the same flat layer as a one-off instruction from yesterday. NanoClaw can't tell the difference, so it either holds onto things too long or drops them at the wrong moment. I ended up building a separate "matters" system to track context somewhat deterministically across sessions (still a work in progress), which tells you something about how unsolved this problem actually is.

Then there's the proactivity dial, which I'm still calibrating. Early on it over-flagged: surfacing things for my attention that didn't need it, checking in when it should have just acted. I fixed that by updating its instruction files to push it toward doing more without asking. That helped, but it introduced a new problem. Duplicate follow-ups to the same person. Tasks showing up twice with slightly different wording. Actions it took that I would've caught if it had just checked in first. You move the dial too far one way and you're drowning in noise; too far the other and it's running ahead of you. The right setting exists somewhere in the middle, and finding it is ongoing.

In closing

Would I keep using it? Yes. But the tool is a work in progress, and so is my setup. The underlying model will improve, the SDK will improve, and I'll keep refining how I run it. The version I'm using in six months won't be the same as the one I started with.

It works in proportion to how much you trust it with. Give it too little and it's just an expensive reminder system. Give it too much and you're cleaning up after it. The interesting zone is somewhere in between, and getting there takes more active management than the demos might suggest.

If you want to try the setup I've been running, it's open source: taslim/nanoclaw-gws-ea. It's a NanoClaw flavour built specifically for Google Workspace — Google Chat, Gmail triage, calendar management, Docs, all running in isolated containers. The instruction files and the matters system are in there too. It's a starting point, not a finished product.