Setting up your spare Mac for Claude Code to control, a step-by-step guide

Original Article ↗ Hacker News Discussion ↗

Enthusiasts are dedicating spare Macs or VMs as always-on “agent” machines for Claude Code, letting the model remotely read, write, and run programs, monitor logs, and handle tasks like triaging alerts or progressing side projects while they’re away from their main computer. Others question whether running AI agents 24/7 is useful or safe, raising concerns about over-automation, security (especially when granting broad system access), and opaque, potentially costly pricing models. The exchange highlights a growing split between those who see autonomous AI agents as powerful new infrastructure for development and operations, and those who view them as overhyped, fragile tools that still require close human supervision.

Use Cases for a Dedicated “Claude Machine”

  • Run long or intensive tasks without tying up a primary laptop: data science jobs, optimization runs, fuzzing, local models (Gemma, Whisper, Qwen), long builds/tests.
  • Keep agents working while away: laptop closed, commuting, or on mobile via SSH, tmux, Remote Control, or iOS apps.
  • Treat the machine as an always-on “sandbox host” for:
    • Log scanning and alert triage (e.g., Datadog, cloud logs, GitHub).
    • Auto-filing issues, simple bug fixes, and test runs.
    • GUI workflows (Chrome, Figma, backup/checklist automation, HomeBridge, home automation).
  • Offload side projects: personal tools, games, content pipelines, system reinstall scripts, etc.

Dispatch / Cowork / Remote Control vs Direct Claude Code

  • Container-based tools (Dispatch/Cowork) are praised for convenience and connectors.
  • Reported limitations: filesystem and network permissions (e.g., PDFs, wget/curl), git lock issues, permissions UX on mobile.
  • Remote Control seen as closer to “direct Claude Code” but still missing some sandbox features.
  • Several users roll their own setups (tmux multiplexing, helper scripts, SSH into homelabs) to work around these gaps.

Hardware Choices and Architecture

  • Many use spare M1/M2/M4 Macs or Mac minis; some reverse it and use a powerful desktop as the headless AI box.
  • Questioning why you need high-end Macs when inference runs in the cloud; proponents argue local tool calls, parallel subagents, and GUI automation benefit from better CPUs/RAM.
  • Alternatives: Ubuntu VMs, libvirt/KVM, Qubes, smol VMs, macOS-in-VM via UTM, or cheap VPSs.

Security and Isolation

  • Opinions range from “give it its own unprivileged user” to stricter isolation:
    • Dedicated VMs, VLANs, firewall deny-all, even joking about Faraday cages.
  • Granting sudo/root is widely called reckless due to prompt injection or escapes.

Skepticism About 24/7 Agents

  • Many commenters struggle to find a compelling “always-on” personal use case beyond log watching or summarizing inboxes.
  • Concerns:
    • Unsupervised agents creating bad or inefficient code.
    • Over-automation adding long-term maintenance burden.
    • AI hype echoing prior bubbles (e.g., NFTs), with “AI bros” building dashboards and content farms.
  • Supporters counter that, when scoped properly, agents offload boring, repetitive work and increase leverage, especially for less tool-proficient developers.

Pricing and Access

  • Anthropic’s per-token pricing seen as complex and anxiety-inducing.
  • Suggested approach: flat-rate subscriptions, treat limits as guardrails, and use one month of heavy usage to calibrate.
  • For financial or identity isolation: ideas include crypto-funded intermediaries, child-like bank accounts, or LLCs with separate accounts.