# How to Run an AI Agent Farm on Old Laptops (The $200 Multi-Agent Setup)

Canonical: https://clawdocx.com/blog/ai-agent-farm-cheap-hardware-macbooks
Author: Sam Okafor
Published: 2026-03-09
Updated: 2026-03-09

> A solo founder in Beijing runs a fleet of AI agents on a cluster of used MacBook Airs for under $200, with the hardware setup explained.

## The Beijing Content Empire

There is a solo entrepreneur in Beijing who manages multiple social media accounts for AI influencers. Not manually — that would be impossible for one person. Instead, they built a fleet of OpenClaw agents that generate content, manage interactions, schedule posts, and handle engagement autonomously.

The hardware? A cluster of second-hand MacBook Air laptops. Each one running its own OpenClaw instance with its own agent. Total hardware cost: less than what most people spend on a single new laptop.

This story, [reported by The Information's Juro Osawa](https://indianexpress.com/article/technology/artificial-intelligence/why-60-year-olds-in-china-are-queuing-up-to-learn-openclaw-10569704/) as part of China's broader OpenClaw adoption wave, illustrates something important: you do not need expensive servers or cloud infrastructure to run multiple AI agents. You need old laptops and good architecture.

## Why Multiple Agents Beat One Big Agent

Before we get into hardware, let us address the obvious question: why not just run one powerful agent?

A single OpenClaw agent is great for personal productivity. But when you are running a business — managing multiple accounts, handling different workflows, serving different purposes — a single agent becomes a bottleneck.

### Isolation

Each agent has its own workspace, its own memory, its own personality. An agent managing a fitness influencer account does not need to share context with an agent managing a tech review account. Isolation keeps things clean and prevents cross-contamination of context.

### Parallelism

AI model APIs have rate limits. One agent can only make so many requests per minute. Five agents across five machines can make five times as many. When you are generating content at scale, this matters.

### Fault Tolerance

If one laptop crashes or an agent gets stuck, the others keep running. You lose one workflow, not everything.

### Specialization

Each agent can be configured with different models, different skills, different SOUL.md personalities. One agent might use Claude for nuanced writing. Another might use a cheap local model for simple scheduling tasks. You optimize per agent instead of compromising on one configuration.

## The Hardware: What Actually Works

Here is the good news: OpenClaw runs on surprisingly modest hardware. The agent itself is lightweight — it is essentially a Node.js process that manages conversations and tool calls. The heavy lifting happens on the AI model provider's servers (unless you are running local models).

### Used MacBook Air (M1 or M2)

- **Cost:** $300-500 used
- **Why it works:** ARM efficiency, silent operation, low power draw (~10W idle), macOS stability, built-in battery as UPS
- **Best for:** Cloud-based model agents (Anthropic, OpenAI) where the laptop just needs to run the gateway and tools

### Used ThinkPad (T480, T490, X1 Carbon)

- **Cost:** $150-300 used
- **Why it works:** Business laptops are built to run 24/7, easy to find used, Linux runs perfectly
- **Best for:** Linux-based setups, developers who prefer Ubuntu/Debian

### Raspberry Pi 5

- **Cost:** $60-100 with case and storage
- **Why it works:** Tiny, silent, 5W power draw, runs OpenClaw fine for cloud model usage
- **Best for:** Single-purpose agents (monitoring, alerts, simple automations)
- **Limitation:** Not enough RAM/CPU for local LLM inference

### Old Desktop PC

- **Cost:** Free (you probably have one)
- **Why it works:** More RAM and CPU than laptops, can run local models if it has a decent GPU
- **Best for:** Local model inference with Ollama, compute-heavy tasks

## Architecture: How to Set It Up

### The Simple Approach: One Agent Per Machine

The easiest setup is one OpenClaw instance per device. Each machine runs independently with its own config, workspace, and model connection.

```
MacBook Air #1  →  Agent: Content Writer  →  Claude Sonnet
MacBook Air #2  →  Agent: Social Manager  →  GPT-4.1 Mini
MacBook Air #3  →  Agent: Research Bot    →  Claude Haiku
ThinkPad #1     →  Agent: Code Reviewer   →  Claude Opus
Raspberry Pi    →  Agent: Monitor/Alerts  →  Local Ollama
```

Each agent connects to its own messaging channels. Maybe one is on Telegram, another is on Discord, another checks email. Or they all report to the same Telegram account through different bots.

### The Coordinated Approach: Hub and Spoke

For more sophisticated setups, designate one machine as the "hub" that runs your primary agent. This agent can delegate tasks to agents on other machines through messaging channels or shared file systems.

```
Hub (MacBook Pro)
  └── Primary Agent (Claude Opus)
        ├── Delegates to Agent #2 (MacBook Air) via Telegram
        ├── Delegates to Agent #3 (ThinkPad) via Discord
        └── Monitors Agent #4 (Pi) via shared log files
```

This mimics how the OpenClaw sub-agent system works, but distributed across physical machines.

### Shared Storage

If your agents need to share files (one generates content, another publishes it), use a shared folder:

- **Syncthing** — Free, open-source, peer-to-peer file sync. Perfect for agent farms.
- **Tailscale + NFS** — If your machines are on different networks, Tailscale creates a mesh VPN and you can mount shared folders.
- **Git repository** — Agents can push/pull content through a shared repo. Surprisingly effective for text-based workflows.

## The Economics

Let us do the math on a realistic setup:

### Hardware (One-Time)

| Item | Cost |
|------|------|
| 3x Used MacBook Air M1 | $1,200 |
| 1x Raspberry Pi 5 | $80 |
| Network switch + cables | $30 |
| **Total** | **$1,310** |

### Monthly Operating Costs

| Item | Cost |
|------|------|
| Electricity (~40W total, 24/7) | ~$4 |
| AI API costs (varies by usage) | $20-200 |
| Internet (you already have this) | $0 |
| **Total** | **$24-204/mo** |

Compare this to cloud alternatives:

- **4 cloud VMs** (even small ones): $40-160/month
- **Managed AI agent hosting**: $50-500/month
- **Hiring humans to do this work**: Do the math yourself

The used hardware pays for itself in months if it replaces cloud costs, and in weeks if it replaces human labor.

## Power Management

Running multiple laptops 24/7 requires some thought about power:

### Keep Them Cool

Laptops in clamshell mode (lid closed) can overheat. Either keep lids open or use a laptop stand that allows airflow. For a cluster of MacBook Airs, a vertical laptop dock rack works well.

### Battery as UPS

One advantage of using laptops over desktops: the built-in battery acts as a UPS. If your power flickers, your agents keep running. On a MacBook Air, the battery can last 4-8 hours under light load.

### Power Monitoring

Use a smart power strip to monitor total power consumption and set up alerts if a machine goes offline. Some smart strips also let you remotely restart outlets — useful for hard-locking issues.

### macOS Settings

If using MacBooks, adjust these settings:

- **Energy Saver** → Prevent automatic sleeping when the display is off
- **Wake for network access** → Enabled
- **Start up automatically after power failure** → Enabled (in System Settings → General → Startup)

## Security Considerations

Running multiple internet-connected agents requires security hygiene:

- **Keep each agent's API keys separate** — If one machine is compromised, the others are not affected
- **Use Tailscale or WireGuard** for inter-machine communication instead of exposing ports
- **Firewall each machine** — Only allow necessary outbound connections
- **Use the [OpenClaw security hardening checklist](/blog/openclaw-security-hardening)** on every machine
- **Physical security** — If these machines are in your apartment, that is fine. If they are in a shared space, lock them down

## Real Use Cases

Beyond the Beijing content farm, here are practical agent farm scenarios:

### Freelance Automation Agency

Run separate agents for each client. Client A's agent monitors their social media. Client B's agent handles their email. Client C's agent generates blog posts. Each agent is isolated, each has its own billing, each can be configured differently.

### E-Commerce Operations

One agent monitors competitor prices. Another manages inventory alerts. A third generates product descriptions. A fourth handles customer inquiry drafts. Four cheap machines, a full operations team.

### Research Pipeline

One agent continuously searches for papers and news. Another summarizes and categorizes findings. A third generates reports. Each focuses on what it does best.

### Family AI Hub

One agent for you (productivity, email, calendar). One for your partner (separate workspace, separate memory). One shared agent in the living room that handles home automation and family scheduling.

## Getting Started Small

You do not need to buy four laptops tomorrow. Start with what you have:

1. **Get one agent running perfectly** on your main machine. Master the workflow.
2. **Find an old laptop** gathering dust in a closet. Install OpenClaw on it with a specialized purpose.
3. **Add machines as you find use cases.** Used MacBooks show up on Facebook Marketplace and eBay constantly.

The point is not to build a server rack. It is to distribute your AI capabilities across cheap, resilient, purpose-built machines — the same way the Beijing entrepreneur did, but adapted to your needs.

---

*Ready to start? [Install OpenClaw for free](/blog/getting-started-openclaw-free), [explore multi-agent orchestration](/blog/multi-agent-teams-automate-everything), or [learn about sub-agents](/blog/openclaw-sub-agents-explained) for software-level parallelism before adding hardware.*