Why GhostClaw
or: how I stopped worrying and gave an AI its own computer
So you use Claude. You're coding all day, you've seen everyone set up OpenClaw, seen the competitors, seen it get bought out by OpenAI. You've tried to use it and it's fucking trash.
The setup is some approximation of the Claude Code terminal. The online settings don't really work. It feels like a step back in time compared to what you're building across Claude Code, Cursor, the rest of it. And worst of all it's 500,000 lines of code - so making changes that don't break everything is non-trivial.
You've done what I've done. Bookmarked every OpenClaw setup guide. Copied every config workaround. Hit bumps at every step and pushed through anyway.
Don't get me wrong - it is a thing of consummate beauty. Watching your agent talk to other agents on forums, self-iterate, spread out into the world, install things, work shit out for itself and get on with it. That's a mind-bendingly impressive workflow that we all know is going beyond the mainstream.
* stares at bookmarks bar, weeps gently *
It's currently like driving an old classic car. One drives in three is fixing it when it breaks down. Which has some charm - gets you seeing how under the hood works, brings joy in itself. But it's not how I want to work when 99% of what I'm doing is strong implementation plans and then having Claude or Codex get on with it and fucking smash it.
It feels like a step back in time. And yeah - my man just made his retirement money from Sam, so he's not gonna iterate this one.
Beautiful. Broken.
500k lines of code. Setup is an approximation. Online settings half-work. Runs wild across your system - which is the point, but it's statistically impossible to set up safely on your main machine.
charming when it works
Elegant. Caged.
400 lines of core code. Completely customisable. Works out-of-the-box. But it runs in containers - great for security, less great if you want your agent to actually do things across your system.
elegant but in a box
If you're choosing between the two and you're already an advanced Claude user, NanoClaw wins. It's not close. What they've built is elegant, beautiful, works out of the box. Adding skills mirrors what you already do in Claude Code, which is a really nice system.
But it lacks some of the esoteric beauty that OpenClaw gives. The full gamut across your system, your files, your chats, the heartbeat, the freedom. I wanted to see if I could build fundamentally NanoClaw with OpenClaw functionality.
I'm setting up a new Mac Mini with no particularly sensitive files. Fresh accounts across the board:
- New iCloud
- New Gmail
- New Vercel
- New GitHub
I really didn't need a complex container or Docker system to protect other files. I want my agent to go wild and show what it can do - but I also want to add skills really easily. Ideally in exactly the same way I already add skills in Claude Code.
NanoClaw's simplicity.
OpenClaw's freedom.
None of the mess.
a frankenstein, if you will
GhostClaw is set up to be its own agent on its own computer, with its own accounts for everything it needs. You interact with it as a co-worker.
A co-worker that knows everything about you. Knows what annoys you, what pleases you, how you like to be talked to. That can read and reply to emails. That you can voice note. That reads your Slack, reads your Notion, adds to both. Gets daily briefings, heartbeat monitoring, full context across everything you're working on.
And runs fucking wild on your computer.
- Containers stripped out - bare metal, full system access
- Telegram groups as standard
- Voice transcription from OpenAI - talk to your agent, hear back
- Security scanning layer on all skill installs
- OpenClaw-style heartbeat monitoring + daily briefings
- OpenClaw-style memory structure recreated
- Full compatibility across the NanoClaw + OpenClaw skill ecosystem
Bought a Mac Mini. Set up a new iCloud account. Set up screen sharing from my MacBook Air to control it. Installed Claude Code in Terminal. Pointed it at GhostClaw. Typed four lines.
$ cd ghostclaw
$ npm install
$ claude
Then /setup-ghostclaw - and Claude Code talks you through everything. Telegram bot from BotFather, WhatsApp group, email if you want it. API keys where needed. About 10 minutes.
Personally I found it useful to have GitHub and Vercel set up too. So I can literally tell my agent to build something, deploy it on the internet, and send me the link.
future is stupid innit
containers
kubernetes
500k lines of code