What Are "Claws"? And Why You Shouldn't Run Them on Your Mac Mini

Andrej Karpathy just posted a mini-essay about buying a Mac Mini to tinker with what he calls "Claws" — persistent AI agent systems that sit on top of LLMs. He names OpenClaw, NanoClaw, zeroclaw, ironclaw, picoclaw. Simon Willison calls "Claw" a term of art for the entire category.

When Karpathy names something, it sticks. He coined "vibe coding." This is the same energy.

Here's his definition:

"Just like LLM agents were a new layer on top of LLMs, Claws are now a new layer on top of LLM agents, taking the orchestration, scheduling, context, tool calls and a kind of persistence to a next level."

I've been building managed infrastructure for exactly this category. Let me break down what Claws are, why running them on a Mac Mini has real tradeoffs, and what the alternative looks like.

What Makes a Claw Different from an Agent?

Regular LLM agents run, do a thing, and stop. You prompt them, they respond, maybe they call a tool, done.

Claws are persistent. They:

Run continuously on hardware or a server
Have their own scheduling — they do things without you asking
Maintain context across sessions and conversations
Communicate via messaging protocols (MCP, etc.)
Orchestrate multiple agents with tool access

Think of it as the difference between running a script and running a service. A Claw is a service. It's always on, always watching, always ready to act.

The Mac Mini Angle

Karpathy bought a Mac Mini specifically to run Claws. The Apple Store told him they're "selling like hotcakes and everyone is confused." Makes sense — decent hardware, small form factor, runs 24/7 at home.

But here's where I have some thoughts, as someone who's been running persistent Elixir services for a while now.

The Self-Hosting Pain List

I love self-hosting. I really do. But running a persistent AI agent system on a box under your desk means you're now responsible for:

Uptime. Your Claw goes down when your power goes out, when your ISP hiccups, when macOS decides it needs to update at 3am. The whole point of a Claw is that it's always on. "Always on except when it isn't" is a rough spec.

Networking. Your Claw needs to talk to the internet — receive webhooks, call APIs, expose endpoints. That means port forwarding, dynamic DNS, TLS certificates, and hoping your router cooperates. Your ISP probably gives you a dynamic IP.

Security. You're running an AI agent with tool access on your home network. It can execute code, make API calls, access file systems. One misconfigured permission and your Claw can see everything on your LAN.

Updates and maintenance. The Claw ecosystem is evolving fast. OpenClaw pushes updates regularly. You need to manage versions, handle breaking changes, keep dependencies current. On a personal Mac Mini, that's manual work.

Process supervision. What happens when a Claw process crashes? On a Mac Mini, it just... dies. You need to build your own restart logic, health checks, and monitoring. This is a solved problem in production infrastructure, but not on your desktop.

Scaling. Today you run one Claw. Tomorrow you want three. Next month you want one per project. A Mac Mini has finite resources and no way to scale horizontally.

Why Managed Hosting Makes Sense

I'm building OpenClawCloud because I think this pain list is going to hit most people who try to run Claws seriously.

The architecture is built on Elixir and runs on Fly.io. Here's why that matters for Claws specifically:

Supervision Trees

Elixir's OTP supervision is designed exactly for this — long-running processes that need to stay alive. If a Claw process crashes, the supervisor restarts it automatically. No cron jobs, no systemd hacking, no Docker restart policies. It's built into the runtime.

Process Isolation

Each tenant's Claw runs in its own isolated process. One Claw crashing doesn't take down another. The BEAM VM was literally built for this — telecom-grade reliability for concurrent, independent processes. Ericsson designed it in the '80s to keep phone switches running. Turns out that's exactly what persistent AI agents need too.

Built-in Scheduling

Claws need to do things on their own schedule. Elixir has Process.send_after, GenServer timers, and libraries like Oban for persistent job scheduling. No external cron needed. The agent's scheduler lives in the same runtime as the agent itself.

Economics That Work

I run 5 Elixir apps on Fly.io for under €50/month total. The infrastructure is efficient enough that hosting multiple Claws per machine is practical without burning through a credit card.

The State of the Ecosystem

Karpathy mentions several projects, each taking a different approach:

OpenClaw — the full-featured option, though Karpathy himself admits he's "a bit sus'd" about running it directly
NanoClaw — ~4000 lines of core code. Karpathy likes that it "fits into both my head and that of AI agents" — auditable and minimal, runs in containers by default
zeroclaw, ironclaw, picoclaw — variations on the theme with different tradeoffs around size, security, and features

The ecosystem hasn't consolidated yet. But the pattern is clear: people want persistent, tool-enabled AI agent systems that run autonomously.

Where This Is Going

Karpathy naming this category matters. "Vibe coding" went from a tweet to a conference talk title in weeks. "Claws" as a term of art is going to follow the same trajectory. Simon Willison is already using it. It even comes with an established emoji: 🦞

The interesting question isn't whether Claws are real — they obviously are. It's whether the infrastructure catches up. Right now, the default path is "buy hardware and figure it out yourself." That works for tinkering. For production use — agents managing your calendar, monitoring your infrastructure, handling customer requests — you need something more robust.

That's the gap I'm building OpenClawCloud to fill. You bring your Claw config, I handle deployment, uptime, and process supervision. No Mac Mini required.

I'm a solo founder building this in Elixir from Braga, Portugal. It's early days, but the foundation is solid — and today, thanks to Karpathy, the category has a name.

If you're running Claws on a Mac Mini and loving it? Respect. That's how all good infrastructure starts — someone tinkering at home until they need more.

I'm @joaosetas on X. Building OpenClawCloud and other Elixir SaaS products in public.