#OpenClaw #Deployment #AI Agent #Mac Mini #AgentPuter #TinyClaw #Tutorial

Four Ways to Deploy OpenClaw: Mac Mini, VPS, Cloud Pod, or Under 60 Seconds

OpenClaw crossed 207,000 GitHub stars this week. Here's how to actually set it up — from bare metal to one-click cloud — without the gaps most tutorials leave in.

@ AgentPuter Lab
$
~ 20 min read

Part 9 of the Agent Infrastructure Series

OpenClaw is the fastest-growing open-source AI project on GitHub right now: 207K stars, 38K forks, 683 contributors, over 12,000 commits. MacStories editor Federico Viticci burned through 180 million tokens with his instance “Navi” and called it “the thing that changed how I use AI.” Mac Mini units sold out at Apple Stores in multiple cities — in part because people wanted a dedicated, always-on machine to run it.

You want to try it. The question is: how do you actually set it up?

The existing tutorials fall into two camps: either they assume you already know what systemd is, or they skip three steps and leave you staring at an error. This guide covers four paths — from building your own local server to deploying in under 60 seconds — so you can pick the one that matches your skill level and priorities.

The four paths:

PathTimeSkill LevelBest For
A. Mac Mini15–30 minTerminal basicsPrivacy-first, full control
B1. Self-managed VPS15–30 minSSH + Linux opsDevelopers, compliance
B2. AgentPuter~2 minBasic CLIMulti-agent, zero DevOps
C. TinyClaw< 1 minNoneEveryone

First: What Is OpenClaw? (30-Second Version)

OpenClaw is not another chatbot you visit in a browser. It’s a personal AI assistant that runs 24/7 in the background and talks to you through the messaging apps you already use — WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage (via BlueBubbles), Microsoft Teams, WebChat, Matrix, and more. Twelve channels and counting.

What makes it different from ChatGPT:

  • Persistent memory. OpenClaw has a SOUL.md file that defines who it is to you — your name, your preferences, your work style, your timezone. It remembers across sessions.
  • Real-world actions. It reads your email, manages your calendar, operates on your files, controls smart home devices, does research, and drafts replies — autonomously.
  • Voice and canvas. Voice Wake and Talk Mode (powered by ElevenLabs) let you speak to it hands-free. Live Canvas (A2UI) gives it a visual workspace.
  • Local-first. Your data stays on your machine. No training on your conversations. No third-party cloud unless you choose one.

What you need to run it:

  • A machine running Node.js 22+ (macOS, Linux, or Windows via WSL2)
  • An AI model subscription: Anthropic API key or Claude Pro/Max OAuth (strongly recommended by the project’s creator), or OpenAI / Google Gemini
  • The official recommendation: Anthropic Pro/Max + Opus 4.6 — best long-context performance and strongest prompt-injection resistance among supported models

Path A: Mac Mini — The Local Server

Who this is for: You have (or want) a Mac Mini. You want 100% local control. You care about data sovereignty. You want iMessage integration and Voice Wake.

Hardware

A Mac Mini M4 base model ($599) is more than enough: 16GB unified memory, 256GB storage. It draws 5–10W at idle — roughly $15/year in electricity. It’s silent, always-on, and natively supports iMessage through BlueBubbles, which no cloud deployment can match.

If you’re buying refurbished, an M2 with 16GB works fine too. Avoid 8GB models — OpenClaw plus the model’s context window will hit swap under heavy use, and swap on SSD shortens drive life.

Prerequisites

Before you start, confirm you have:

  • macOS 13 (Ventura) or newer
  • Homebrew installed (/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)")
  • Node.js 22+ (brew install node@22, then verify with node --version)
  • An Anthropic API key, or a Claude Pro ($20/mo) / Max ($100/mo) subscription for OAuth login
  • A Telegram account (easiest channel to start with)

Step-by-Step Installation

Step 1: Install OpenClaw globally

npm install -g openclaw@latest

You can also use pnpm (pnpm add -g openclaw@latest) or bun. All three are officially supported.

You should see output ending with something like:

added 1 package in 12s

Verify the installation:

openclaw --version

Expected output: 2026.2.17 (or the latest release number).

Step 2: Run the onboarding wizard

openclaw onboard --install-daemon

The wizard walks you through everything interactively:

  1. Choose your model. Select Anthropic Claude (recommended). If you have a Claude Pro/Max subscription, choose OAuth login — no API key management needed.
  2. Enter your API key or authenticate via OAuth.
  3. Choose a messaging channel. Telegram is the simplest to start with. The wizard will tell you to message @BotFather on Telegram, create a new bot with /newbot, and paste the token back.
  4. Install the background daemon. On macOS, this creates a launchd service so OpenClaw survives reboots and runs even when the terminal is closed.

The whole process takes about 5 minutes.

Step 3: Verify the installation

openclaw doctor

This is the official diagnostic tool. It checks your Node.js version, Gateway connectivity, channel configuration, model access, and daemon status. Every check should show a green checkmark. If something is yellow or red, the doctor will tell you exactly what to fix.

Step 4: Pair your first device

Open Telegram and send any message to your new bot. The bot will reply with a pairing code — this is OpenClaw’s default security mechanism. Unknown senders get a code; they can’t interact with your assistant until you approve them.

openclaw pairing approve telegram <CODE>

That’s it. Your bot is now live. Send it “What can you do?” and watch it respond.

Troubleshooting: If the bot doesn’t respond, run openclaw doctor first. Common issues: the daemon didn’t start (run openclaw onboard --install-daemon again), or the Telegram token has a typo (check ~/.openclaw/openclaw.json).

After Installation

Three things to do immediately:

  1. Write your SOUL.md. Open ~/.openclaw/workspace/SOUL.md in any text editor. Tell it your name, what you do, your communication preferences, your timezone. This isn’t a prompt — it’s an identity file that persists across every conversation.

  2. Connect your tools. Gmail (via Pub/Sub), Google Calendar, Notion, Todoist, Slack — OpenClaw connects through MCP tools and Skills. The wizard may have already prompted you for some of these.

  3. Run a real task. Don’t start with “tell me a joke.” Try: “Summarize my unread emails and sort them by priority” or “What’s on my calendar tomorrow and do I have any conflicts?”

Real Costs

ItemCost
Hardware$599 one-time (Mac Mini M4 base)
API (moderate use)$15–50/month (or $20/mo with Claude Pro subscription)
API (heavy use)$100–300/month (Viticci-level)
Electricity~$15/year
Setup time15–30 minutes

Path B: Cloud Deployment — Self-Managed or AgentPuter

Who this is for: You don’t have a Mac. You use Linux or Windows. You want to access your assistant from anywhere. You need real 24/7 uptime, not “my laptop is open.”

There are two approaches to cloud deployment: manage your own server (full control, full responsibility) or use AgentPuter (a cloud runtime built specifically for AI agents, zero DevOps). We’ll cover both.

B1: Self-Managed VPS

Who this is for: You know SSH. You want root access. Your company requires self-hosted infrastructure. You’re optimizing for cost.

Choose a Server

ProviderMinimum SpecMonthly Cost
Hetzner2 vCPU, 4GB RAM, 20GB SSD~$5/mo
DigitalOcean2 vCPU, 4GB RAM, 25GB SSD~$12/mo
Vultr2 vCPU, 4GB RAM, 25GB SSD~$12/mo

Warning: Don’t go below 4GB RAM. With Docker, 2GB will OOM-kill the process. Trust me on this one.

OS: Ubuntu 22.04 LTS.

Windows users: You don’t need a VPS. OpenClaw officially supports Windows via WSL2, and the README marks it as “strongly recommended.” Install WSL2, then follow the same Linux instructions below.

ssh root@your-server-ip

# Install Node.js 22
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt-get install -y nodejs

# Install OpenClaw
npm install -g openclaw@latest

# Run the wizard
openclaw onboard --install-daemon

On Linux, the daemon installs as a systemd user service instead of launchd. Everything else is identical to the Mac Mini path — the wizard handles it.

Verify:

openclaw doctor

Option B: Docker Compose (Production-Grade)

git clone https://github.com/openclaw/openclaw.git
cd openclaw
./docker-setup.sh

Advantages:

  • Process isolation — OpenClaw runs in its own container
  • One-command restart: docker compose restart
  • Built-in log management
  • Sandbox mode: set sandbox.mode: "non-main" to run group/channel sessions in isolated Docker containers

Disadvantages:

  • Requires Docker Engine 24+
  • Another layer of abstraction to debug

The Gateway dashboard is available at http://localhost:18789 — but only locally, which brings us to the most important part.

Remote Access: Do This Right

This is not optional. CVE-2026-25253 demonstrated that exposing the Gateway port directly to the internet enables token exfiltration leading to remote code execution. Don’t do it.

Recommended: Tailscale Serve/Funnel

Tailscale gives you a private network overlay. OpenClaw has first-class support:

{
  "gateway": {
    "tailscale": {
      "mode": "serve"
    }
  }
}
  • "serve" — accessible only within your Tailscale network (most secure)
  • "funnel" — public HTTPS, but requires password authentication (gateway.auth.mode: "password")

Alternative: SSH tunnel

ssh -L 18789:localhost:18789 user@your-server-ip

Then access the dashboard at http://localhost:18789 on your local machine.

Backup

Your entire OpenClaw state lives in ~/.openclaw/:

PathContents
openclaw.jsonMain configuration
workspace/SOUL.md, AGENTS.md, TOOLS.md, skills/
credentials/Channel credentials (WhatsApp session, Telegram token, etc.)

Back this up daily. A simple cron job works:

tar czf ~/openclaw-backup-$(date +%Y%m%d).tar.gz ~/.openclaw/

The Pain Points of Self-Managing

I’ve been running OpenClaw on a self-managed VPS for three months. Here’s what I’ve learned:

  • You have to install Node.js, configure systemd, set up Tailscale, and handle TLS yourself.
  • OAuth tokens expire. If you’re sleeping in a different timezone, your WhatsApp connection dies at 3 AM and your agent goes silent until you wake up and re-authenticate.
  • After a server reboot (kernel update, provider maintenance), OpenClaw doesn’t always come back cleanly. You learn to write recovery scripts.
  • Running multiple agents requires manual process isolation and resource allocation.

In one sentence: a self-managed VPS gives you maximum freedom, but you’re also your own DevOps team.

If that last paragraph made you tired, there’s another way.

B2: AgentPuter — A Cloud Runtime Built for AI Agents

Who this is for: You want cloud deployment without the DevOps. You need multi-agent support. You care about 24/7 reliability with SLA guarantees.

What Is AgentPuter?

AgentPuter is not a generic VPS. It’s a dedicated cloud runtime designed specifically for AI agents — OpenClaw, ClawBot, MoltBot, and custom agents. Think of it as “Heroku for AI assistants”: you don’t manage the server, the container, or the daemon. You get a Pod that never shuts down.

It solves the four pain points of self-managed VPS:

Self-Managed VPS ProblemHow AgentPuter Solves It
Server reboots = agent offlinePod runs 24/7 with auto-recovery and uptime SLA
OAuth tokens expire = manual re-loginServer-side token management with automatic refresh
Agents consume local CPU/memoryCloud-isolated resources; your local machine is unaffected
Access only from same networkControl your agents from any device, anywhere

Three-Step Deployment

# Step 1: Create your Pod
agentputer create openclaw
# → Allocates a cloud environment, choose your config

# Step 2: Connect your services
agentputer connect
# → Authorize Google, Notion, Slack, etc. Credentials stored securely in-pod

# Step 3: Deploy OpenClaw
agentputer deploy openclaw
# → Your agent starts working 24/7. WhatsApp, Telegram, Discord — all channels live

Total time: about 2 minutes.

Why AgentPuter Stands Out

  • Multi-agent parallel execution. Run ClawBot (personal assistant) + MoltBot (calendar/scheduling) + a custom research agent in the same Pod, each with isolated resources.
  • Auth that never expires. OAuth tokens are maintained server-side with automatic refresh. Your WhatsApp connection won’t die at 3 AM because a token expired.
  • Access from anywhere. Phone, tablet, another computer — your AI follows you, not the other way around.
  • Current status: Early Access. Request an invite code at agentputer.com — codes are sent within 24 hours.

AgentPuter vs. Self-Managed VPS

Self-Managed VPSAgentPuter
Deploy time15–30 minutes~2 minutes
OperationsYou manage systemd, Tailscale, backup, TLSZero ops — Pod is auto-maintained
Multi-agentManual isolationNative parallel support
Auth managementHandle token expiry manuallyAutomatic refresh
Monthly costVPS $5–24 + API feesPod subscription + API fees
Control100% root accessLimited to Pod environment
Best forStrong ops skills / complianceWant simplicity / multi-agent users

Path C: TinyClaw — 60 Seconds, Zero Terminal

Who this is for: You don’t want to learn command-line tools. You don’t want to manage servers. You just want to use an AI assistant, not install one. You want to try it right now.

Why This Path Exists

Path A requires $599 and knowing what Node.js is. Path B (self-managed) requires SSH, Docker, and Tailscale. Even AgentPuter, while much simpler, still involves a CLI and an invite code.

Most people see npm install -g and close the tab.

OpenClaw’s 207,000 stars prove the demand is real. The bottleneck isn’t the product — it’s the deployment barrier. TinyClaw eliminates it.

What Is TinyClaw?

TinyClaw is the consumer-grade product built by the AgentPuter team. If AgentPuter is “the developer’s AI agent cloud,” TinyClaw is “everyone’s AI agent entry point.”

Three steps to go live:

  1. Choose your model — Claude Opus 4.6, GPT-5.2, or Gemini 3
  2. Choose your channel — Telegram (Discord and WhatsApp coming soon)
  3. Sign in with Google → deployment complete

No server. No SSH. No Node.js. No terminal. The infrastructure is pre-configured and waiting to be assigned to you.

From click to first conversation with your AI assistant: under 60 seconds.

What You Can Do With It

Once deployed, your TinyClaw-hosted OpenClaw can:

  • Read, summarize, and draft replies to your emails
  • Manage your calendar, set reminders, resolve scheduling conflicts
  • Summarize documents, draft contracts, generate invoices
  • Track expenses, compare prices, assist with tax preparation
  • Conduct competitive research, draft social media posts, track OKRs
  • Monitor news feeds, book travel
  • Learn new capabilities through natural language — just tell it what you need

The Ultimate Comparison

Mac MiniSelf-Managed VPSAgentPuterTinyClaw
Deploy time15–30 min15–30 min~2 min< 1 min
Technical skillTerminalSSH + opsBasic CLINone (GUI only)
Hardware cost$599$0$0$0
Monthly costAPI onlyVPS + APIPod + APIHosting + API
Data location100% localYour VPSAgentPuter cloudTinyClaw cloud
24/7 reliabilityDepends on your MacDepends on your opsSLA-backedSLA-backed
Multi-agentManual configManual isolationNative parallelSingle agent
Auth managementManualManualAuto-refreshAutomatic
iMessageYes (BlueBubbles)NoNoNo
Voice WakeYes (macOS app)NoNoNo
WindowsNoYes (WSL2)Yes (CLI)Yes (browser)
Best forPrivacy / power usersDevOps / complianceDevelopers / multi-agentEveryone

After Setup: 7 Steps to Turn Your OpenClaw from Toy to Employee

Regardless of which path you chose, installation is just the beginning. Here’s how to make it actually useful:

1. Write Your SOUL.md

Location: ~/.openclaw/workspace/SOUL.md

This is not a system prompt. It’s an identity file. Tell it:

  • Your name, job title, and what you do
  • Your communication style (formal? casual? bullet points?)
  • Your timezone and working hours
  • Your preferences (“I prefer Markdown,” “Never schedule meetings before 10 AM”)

The more specific you are, the less you’ll need to repeat yourself.

2. Run openclaw doctor

Do this after setup, after every upgrade, and whenever something feels off. It checks everything — Node version, Gateway health, channel connectivity, model access, daemon status — and tells you exactly what to fix.

3. Learn the Chat Commands

These work in any connected channel — Telegram, WhatsApp, Slack, Discord:

CommandWhat It Does
/statusShows current model, token usage, session info
/new or /resetResets the conversation session
/compactCompresses context to save tokens
/think highEnables deep thinking mode (Opus 4.6)
/verbose onMore detailed responses
/usage fullShows token consumption after each reply

These are your daily controls. /compact alone can save you 30–40% on token costs in long conversations.

4. Connect Your Tools

OpenClaw uses MCP (Model Context Protocol) tools and Skills to integrate with external services:

  • Gmail — via Pub/Sub for real-time email notifications
  • Google Calendar — read, create, and modify events
  • Notion / Todoist / Linear — task management
  • Slack — both as a channel and as a tool
  • Browser — OpenClaw can control a dedicated Chrome instance

5. Plant Memory Seeds

Give it initial context that it will remember forever:

“I have a team meeting every Monday at 10 AM. My manager’s name is Sarah. I prefer responses in Markdown. I’m working on the Q1 launch for Project Atlas.”

These facts persist in memory and inform every future interaction.

6. Run a Real Workflow

Don’t test with trivia. Give it work:

  • “Summarize my unread emails and sort them by priority”
  • “What’s on my calendar this week? Flag any conflicts”
  • “Research the top 5 competitors for [product] and give me a comparison table”
  • “Draft a reply to the email from [person] — professional tone, accept the meeting but suggest Thursday instead”

7. Install Community Skills

Enable ClawHub in your config and your agent can automatically search for and install new Skills as needed. You can also browse manually at SkillsMP — the marketplace has thousands of community-contributed Skills for everything from Jira integration to flight price tracking.

FAQ

How much does the API cost? Light use: ~$15/month. Moderate: $30–50/month. Heavy (Viticci-level): $100–300/month. Alternatively, a Claude Pro subscription ($20/month) or Max ($100/month) lets you authenticate via OAuth without managing API keys separately.

Which model is best? The project creator strongly recommends Claude Opus 4.6 — it has the best long-context performance and the strongest resistance to prompt injection. GPT-5.2 and Gemini 3 are also supported. Pick whichever you already have a subscription for.

Is it secure? By default, OpenClaw uses DM pairing — any unknown sender receives a pairing code and cannot interact with your assistant until you approve them with openclaw pairing approve. Local deployments keep all data on your machine. For remote access, use Tailscale — never expose the Gateway port (18789) directly to the internet.

Does it support Chinese? Yes. The underlying models support Chinese natively. There’s a community documentation site at clawd.org.cn, and it works with domestic models like DeepSeek, Moonshot Kimi, and Qwen.

Does it work on Windows? Yes, via WSL2 (Windows Subsystem for Linux) — officially supported and strongly recommended. Or use TinyClaw for a browser-based experience with zero local setup.

How is this different from ChatGPT Plus? ChatGPT waits for you to visit it. OpenClaw reaches out to you — through WhatsApp, Telegram, Slack, wherever you are. ChatGPT has no persistent memory across sessions; OpenClaw does. ChatGPT can’t operate on your files or email; OpenClaw can. ChatGPT doesn’t support Voice Wake or Live Canvas; OpenClaw does.

Something went wrong. Now what? Run openclaw doctor. It auto-diagnoses and gives targeted fix instructions. The community Discord has 5,000+ active members who can help with edge cases.

How do I update?

openclaw update --channel stable

There are also beta and dev channels if you want cutting-edge features.

Closing Thoughts

Four paths, one destination: your own 24/7 AI assistant.

  • Mac Mini if you’re a power user who wants full control, Voice Wake, and iMessage.
  • Self-managed VPS if you’re a developer who wants root access and maximum flexibility.
  • AgentPuter if you want cloud reliability, multi-agent support, and zero DevOps.
  • TinyClaw if you just want it to work — 60 seconds, no terminal.

These paths aren’t mutually exclusive. You can start with TinyClaw to experience OpenClaw in under a minute, graduate to AgentPuter when you want to run multiple agents, and eventually build a fully sovereign setup on a Mac Mini when you’re ready for complete control.

This is Part 9 of our Agent Infrastructure series. We’ve covered why your agent needs its own computer, OpenClaw’s architecture, skills ecosystem, enterprise workflows, the ClawdBot deep dive, business models, and what it means when its creator joins OpenAI. Today was the “do it yourself” chapter.

If you’ve deployed successfully, come tell us in the comments or on Discord: what did you name your agent, and what was the first real task you gave it?

References: