Build an AI Agent Team with OpenClaw

A dedicated machine sitting in a home office. Four AI agents running on it. Each with a name, a role, a personality — and a queue of real work to do for my client's business.

That's the setup I recently built for a client, Steve. In this post, I break down exactly how I did it: hardware choices, security setup, multi-agent config, cost management, and what we're actually using these agents for.

If you're wondering whether a team of AI agents is worth the effort — and how to set one up for a real business — this is the guide.

What Is OpenClaw and How Is It Different from Claude Code?

OpenClaw (previously called Claudebot, then Moltbot) is an always-on AI agent framework. At its core is a process called the gateway — a server running on a dedicated machine that can use tools, run bash scripts, browse the web, and more. Unlike other tools, these agents run on their own machine, 24/7, ready to pick up tasks whenever you need them.

Claude Code can do a lot of that too. So what's the difference?

Claude Code lives in your terminal. It only works while you're actively running a session. When you close it, it stops.

OpenClaw is always running. It maintains a persistent workspace with memory and session logs. You delegate tasks via WhatsApp, Slack, or Telegram and walk away — they do the work, and you check in when you want. They handle it in the background, on their own machine, while you're sleeping or playing with your kids.

That's a fundamentally different paradigm. Instead of you managing an AI tool, you have AI teammates with their own workstations.

Where Should You Run OpenClaw? VPS vs. Dedicated Machine

The first rule: don't run OpenClaw on your personal machine.

You don't want to give it access to all your files and accounts. And your laptop can't be on 24/7 anyway. OpenClaw needs its own home.

You have two options:

Option 1: Cloud VPS (~$5–20/month)

• Fast to set up
• Cheap to start
• Always on, globally accessible
• Works great for most people

Option 2: Dedicated Physical Machine

• Screen sharing access — see the actual desktop
• More storage and bandwidth at no extra recurring cost
• SSH access for quick commands
• Higher upfront cost, lower ongoing cost at scale

For Steve, I recommended a Mac Mini M4 (~$600). My reasoning: we wanted to be able to screen share in visually to manage the environment like a real workstation, and we needed more storage than the entry-level VPS tiers usually offer without getting hit with high recurring monthly fees. Plus, having local bandwidth for video processing tasks was a huge factor.

How to Secure Your AI Agents: The Employee Onboarding Approach

This is where most people underestimate the work involved. Setting up the machine is the easy part. Setting up access correctly is where you need to think carefully.

My approach for Steve was to treat every agent like a new employee. That means:

• Dedicated email addresses — agents get their own email, not access to the client's inbox
• Separate GitHub accounts — I invite them to specific repos only, not the whole organization
• Scoped file access — agents see what they need, nothing more
• Controlled Cloud Storage — I set up separate accounts for the agents, then share only specific folders from Steve's main account to the agent accounts

The storage trick is essential. Steve's main account has all his personal and business files. We don't want agents touching any of that. So I created a secondary "Agent Storage" account for OpenClaw, and only bridged the specific folders that the agents actually need to work in.

Managing Costs: OpenRouter, Model Selection, and Token Budgets

Here's the real talk on costs: without monitoring, it can add up fast.

During the initial setup and "training" phase for Steve's agents, we hit nearly $200 in token costs within the first two days. This was mostly from chatting, configuring, and testing the workflows — not even full production use yet.

The Claude Max Plan Question

Many clients already pay for Claude Max ($100–200/month). Can we just use that for their OpenClaw agents?

My advice: Keep them separate. Anthropic's terms of service are ambiguous on heavy automation on personal plans. The safer path I take for my clients is to keep their personal Claude Max plan for their own manual use and route all OpenClaw agents through separate API tokens.

Why I Use OpenRouter

Instead of managing separate API keys for Anthropic, OpenAI, and others, I route all of Steve's agent API calls through OpenRouter. Benefits:

• One dashboard for all client usage and costs
• Easy model switching without changing the code
• Access to hundreds of models across providers
• I can route expensive tasks to powerful models, and cheap tasks to fast/cheap models automatically

What Chat Interface Works Best for OpenClaw? WhatsApp vs. Slack

OpenClaw supports many chat tools. I initially tested Telegram for Steve — it's the easiest to set up. But we quickly hit a friction point: markdown rendering.

When agents send back formatted responses (code blocks, tables, bullet lists), Telegram's rendering can be messy. It kind of works — but it's not professional.

Most of my business clients already use Slack or WhatsApp. For Steve, we primarily use WhatsApp for quick delegation and Slack for deeper work. Slack has been the winner for complex output because of:

• Excellent markdown support — code blocks render perfectly
• Threaded replies — Steve can manage multiple agent conversations without losing the plot
• One Slack bot per agent — clean separation, easy to identify who's responding
• Familiar UX — it feels like chatting with a real team member

Multi-Agent Configuration: How I Set Up the Team for Steve

Here's the setup I landed on for Steve's business — four agents, each with a distinct role, name, and model assignment.

How to Configure Multiple Agents in identity.md

OpenClaw has a single identity.md file that normally defines one agent. For Steve, I used it to define all four — one section per agent. I gave them distinct personalities so Steve knows exactly who he's talking to.

# Agent Identities

## Rocky — System Admin
Role: System administrator for OpenClaw infrastructure
Model: claude-opus-4-5 (via OpenRouter)
Focus: Config, system health, infrastructure tasks

## Dev-Agent — Developer
Role: Software developer
Model: claude-opus-4-5 (via OpenRouter)
Focus: Backlog issues, PRs, production errors

## Marketing-Agent — Marketer
Role: Marketing and content
Model: claude-sonnet-4-6 (via OpenRouter)
Focus: Content pipelines, social media, newsletters

## Assistant-Agent — General Assistant
Role: General purpose assistant
Model: claude-sonnet-4-6 (via OpenRouter)
Focus: Glue work, scheduling, documentation, admin

Shared Workspace vs. Isolated Workspaces

I chose to have all of Steve's agents share one workspace. This means they all access the same memory files, configurations, and AGENTS.md directives.

The advantage: shared context. When the Dev-Agent makes a change, the Marketing-Agent can see what shipped. When the Assistant-Agent documents a new process, it's available to everyone on the team.

The tradeoff: less isolation. For Steve's business, the shared approach works perfectly. For larger organizations with strict data separation needs, I would set up separate workspaces.

Sub-Agents for Task Delegation

Each top-level agent can spawn sub-agents when needed. This is how I handle cost control for Steve: if the Assistant-Agent (running on Sonnet) hits a task that needs real reasoning power, it delegates to a sub-agent running on Opus — rather than switching models mid-session.

What Should Your AI Agents Actually Do? Real Use Cases

This is the most important question — and the one that trips up most people. "Having an AI agent" isn't useful. Having one with a specific, recurring job is.

I identified four areas where agents fill real gaps in Steve's business:

1. Content Pipeline

Steve has a lot of "raw" knowledge that usually stays in his head or in rough meeting notes. We set up an agent to observe his shared folders, capture these ideas, and surface them as drafted blog posts or social updates. Less good stuff slips through the cracks.

2. Development Backlog

Steve isn't a full-time dev, but he maintains several internal tools. The Dev-Agent picks up backlog issues, tracks errors, and submits PRs. The work keeps moving even when Steve is busy with clients.

3. Glue Work

Every minute spent project managing, copying and pasting data, or updating documentation is a minute not spent growing the business. That's the Assistant-Agent's domain. It's the "small" work that is a constant drain on a founder's time.

4. Reporting & Pattern Recognition

This is my favorite use case: having agents surface trends Steve wouldn't spot. We have agents watching the business data and generating regular reports on anomalies and growth opportunities. It's a new level of operational awareness.

Building a Custom HQ Dashboard for the Team

OpenClaw has a built-in cron system for scheduled tasks — but it doesn't associate tasks with specific agents cleanly out of the box. To give Steve full visibility, I built him a custom HQ dashboard.

It's a simple web app that connects to the OpenClaw gateway and shows him:

• A list of all scheduled tasks with agent assignments
• Token usage tracking per agent in real-time
• A central view of everything running in his system

I built this for him in about a day using Claude Code. The lesson: your AI agent system might need custom tooling to truly fit your workflow, and building it quickly is part of the process.

Is OpenClaw Worth It? My Honest Take

My honest take: OpenClaw is still early and can be raw. I spent more time than expected getting Steve's initial configuration perfect. If you go in expecting "magic" without any setup, you'll be frustrated.

But the signal is undeniable.

What's Working

• The multi-agent paradigm is a game-changer for solo founders
• Slack and WhatsApp integration is clean and practical for real work
• Shared workspace memory means agents get better at their jobs over time
• The "teammate" model maps much better to business needs than a standard chatbot

What's Still Raw

• Built-in cron scheduling needs more polish
• Cost management requires active attention
• The tool is evolving fast — you need to be ready to adapt

Step-by-Step: Setting Up Your First Multi-Agent OpenClaw Team

Step 1: Provision a Dedicated Machine

Spin up a VPS (Digital Ocean, Hetzner) or a dedicated physical machine. Install Node.js and follow my OpenClaw setup guide.

# Install OpenClaw globally
npm install -g openclaw

# Initialize your workspace
openclaw gateway init

# Start the gateway
openclaw gateway start

Step 2: Set Up Security Boundaries

• Create a dedicated email for your agents
• Create a separate GitHub account with limited repo access
• Set up secondary cloud storage accounts for agent file access
• Share only the specific folders agents need

Step 3: Configure OpenRouter

Sign up at openrouter.ai, generate an API key, and configure it. This becomes the single "wallet" for your agent team.

# In your OpenClaw config, set the model provider to OpenRouter
# Example: claude-sonnet-4-6 via OpenRouter
# openrouter/anthropic/claude-sonnet-4-6

Step 4: Create Agent Identities

Edit your identity.md file to define your team. Give each agent a name, role, and default model.

Step 5: Set Up Messaging Channels

1. Connect your WhatsApp account via the OpenClaw CLI
2. Create Slack apps for each of your agents at api.slack.com/apps
3. Install them to your workspace and add the tokens to OpenClaw

Step 6: Assign Initial Tasks

Start small. Give one agent one recurring job. Run it for a week, evaluate the results, and expand from there.

# Example: Daily morning briefing task
Agent: Assistant-Agent
Task: Generate a morning briefing for me on WhatsApp with:
  - 3 priority tasks for today
  - Any unread business messages
  - A summary of production logs from the night
Schedule: Daily at 8:00 AM
Model: claude-sonnet-4-6

Key Takeaways

• OpenClaw ≠ Claude Code — it's always-on, persistent, and runs on its own machine
• Security first — treat agents like employees: scoped access, dedicated accounts, limited permissions
• Costs are real — use OpenRouter and model routing to control spend
• WhatsApp & Slack are the best channels for daily use and complex output
• Multi-agent via identity.md — define multiple agents, assign models based on task type
• Start with one job per agent — prove value before expanding scope
• Custom dashboards — build what you need to manage your team effectively
• Early adopter advantage — this paradigm is here to stay; now is the time to build the skills

FAQ

Can I run OpenClaw on Windows?

OpenClaw is designed for Unix-like environments (macOS, Linux). If you're on Windows, running it on a Linux VPS is the recommended path.

What's the minimum VPS spec for running OpenClaw with multiple agents?

A 2 vCPU / 4GB RAM VPS is a solid starting point. If you plan to run 4+ agents with parallel tasks, consider 4 vCPU / 8GB RAM. Monitor CPU and memory as you scale up.

Do I need a paid Claude plan to use OpenClaw?

No. OpenClaw uses API tokens, not your Claude subscription. You pay directly for API usage — either via Anthropic directly or through OpenRouter. Keep your personal Claude plan (if you have one) separate.

How do agents share memory in OpenClaw?

When agents share a workspace, they access the same set of markdown memory files. Any agent can read or write to these files, which means shared context builds up over time. You control what goes in the memory through your AGENTS.md and workspace structure.

What happens if an agent makes a mistake?

This is why scoped access matters. If an agent only has access to specific repos, specific folders, and specific services — a mistake is contained. Always test agents in low-stakes environments before giving them access to production systems.