GSV: A Mind for Your Machines
So you don’t have to look around: https://github.com/deathbyknowledge/gsv and gsv.space
For the last year I’ve been thinking about the trends in human computer interfaces and what the next big step is. And I don’t mean how we went from CLIs to GUIs.
It sort of went like this:
- The computer was as big as your living room. The interface was a terminal connected to it, single use.
- Machine and software get better, now multiple terminals using the same big computer are available.
- Hardware gets small enough you can have a computer at home. This begins personal computing and it fits on your office desk.
- Fast forward 20 years or so. You now have a phone which is basically a digital extension of yourself (and fits in your pocket).
- You might think they can’t get smaller. Well, they got so small that a lot of computers moved to the cloud.
Even as this happened, you didn’t stick with a single device. It’s not like you have a phone or a laptop. Your digital life is fragmented between devices that serve different purposes.
Ok so… what’s the next big thing? Take a wild guess.
The Only Way To Win Is If Everybody Has One
Fuck, enshittification is real. Ads, spam, engagement bait, slop, etc.
Shit, it’s so bad that even the simplest social platform, EMAIL, is now garbage. Only 1 out of 10 emails in my inbox are relevant to me.
We need a cognitive firewall. This is why I think personal AI is the answer: a cracked computer user with access to all your devices, aligned to YOU.
I’m not fully there yet but…
GSV is my first step.
A year ago I started ideating about smarter computers but I was very early and still had a lot to learn on how to use this new technology. Fast-forward many ideas (or just look at my GitHub) and the release of OpenClaw, which very much helped me piece the last few things together, and made me finalize GSV.
It is not “yet another efficient OpenClaw”. I am not going for feature parity and I’ve already deviated in how they work, yet they have similar features.
- Natively on Cloudflare: The control plane, file storage and basically all infrastructure are Cloudflare Primitives. (shoutout to my fellow DO enjoyers)
- Distributed by design: All your machines available to your agent. Globally available without exposing a single port.
- Async Agents: The LLM calls happen in an isolate in some Cloudflare datacenter, but the tool calling happens in your local devices.
Why Cloudflare?
I want open and local AI but damn infrastructure is annoying. Cloudflare makes this really easy and it’s the best fit for this technology. If there was something better I would’ve used it instead.
This is why I’ve tried making it as easy as possible for you to deploy your own and I’ve taken special care so that it all fits on a Free tier Cloudflare account. Again, open for everyone
Isn’t this Moltworker?
Don’t confuse GSV with Moltworker, which is more or less a wrapper around OpenClaw that runs 24/7 in a Cloudflare Sandbox and is exposed through a Worker. The monthly cost of the Sandbox is ~34$.
GSV is an implementation from scratch. It’s designed to be distributed, run on Cloudflare Workers directly and doesn’t use Sandboxes nor Containers. Even if you’ve never heard of Cloudflare before, you can create an account today and use GSV for 0$ a month.
Repeat the distributed part please

Ok so, when you finish your GSV deployment, you’ll have a few components:
- Gateway: A Durable Object (i.e. Worker with persistent storage) that routes traffic from Clients, Channels and Nodes, manages configuration and makes everything work together.
- Clients: How you talk to your agent. Right now that’s WhatsApp, Discord, a CLI or WebUI but the point is you can reach it from anywhere. Your agent is always on, even when your devices aren’t.
- Channels: The messaging integrations (WhatsApp, Discord, etc.). These connect to the Gateway and relay messages both ways.
- Nodes: Your actual machines. Your laptop, your home server, your Raspberry Pi or whatever you want your agent to have access to. Each one runs a lightweight CLI process that connects to the Gateway over WebSocket. No ports exposed, no tunneling or VPNs. The node reaches out to Cloudflare, not the other way around.
Your agent lives on the edge, but acts on your machines. The LLM reasoning happens in a Cloudflare isolate (fast, global and cheap), while tool execution (reading files, running commands, etc.) happens locally on whichever node has the right capabilities. Your agent can cat a log on your server and git push on your laptop in the same conversation.
And because nodes connect outbound, your devices can be behind NAT, on different networks, in different countries. Doesn’t matter. They all show up as available tools to one unified agent.

It’s early
This has been the personal project that I’ve used the most. I basically can’t stop. But it’s early, there is a lot of stuff I want to add and get right.
I hate hype and if you came hoping to stir the pot, please don’t and read this.
Now, if you’re fine hacking around or simply appreciate cool technology that might have a bug or two, please try it out for yourself and let me know how it goes.