I've been running OpenClaw for about a month now. Started with one agent, ended up with eight. Here's how the setup works and what it actually costs.

The ProblemThe Problem

One AI assistant hits a wall fast. It can't be a coder AND a project manager AND handle social media — at least not well. Context gets bloated, responses slow down, and you spend more time re-explaining things than getting work done.

So I split it up.

8 Agents, 4 Nodes8 Agents, 4 Nodes

Each agent has a name, a personality, a Discord bot account, and a specific job:

• Patoo — main brain, orchestrator (runs on Claude Opus)
• Vandana — dev/engineering (Claude Sonnet)
• Naomi — project management (Sonnet)
• Calypso — executive assistant (Sonnet)
• Sparks — social media / marketing (Sonnet)
• Jake — support (Sonnet)
• Corazón — general assistant (Sonnet)
• Lori — UI/UX design (Sonnet)

Only the orchestrator needs the expensive model. Everyone else runs on Sonnet — same quality for their specific tasks, way cheaper on token usage.

They talk to each other in a shared Discord channel. Tag-based routing — Naomi can tag Vandana to ask about a PR, Sparks can tag Patoo for content approval. Each agent only responds when mentioned, so they're not burning tokens on every message.

The HardwareThe Hardware

Here's where it gets interesting. I'm running this across a mix of what I already had plus a $599 Mac Mini:

The OpenClaw gateway runs on MacMax. It handles all the AI model calls, session management, and message routing. The other nodes are just exec targets — when Vandana needs to run a shell command or compile something, it gets routed to ForgeMini via SSH tunnel. The gateway still does the thinking.

What It CostsWhat It Costs

Here's the actual monthly breakdown:

AI Subscriptions:

• Claude Max 20x plan: $200 (all 8 agents share this)
• OpenAI Plus: $20 (fallback models)
• ElevenLabs Creator: $22 (TTS)

Infrastructure:

• DigitalOcean VPS: $72 (two droplets)
• MacMax electricity: $3.46 (40W avg × 24h × 30d × $0.12/kWh)
• ForgeMini electricity: $0.86 (10W avg)

Total: $318.32/month

Once I finish migrating off the vandana-dev VPS, that drops to $270.

The Claude Max plan is the big one. But it covers unlimited usage for all 8 agents — no per-token anxiety. The fallback chain goes Anthropic → OpenAI GPT-5.2 → GPT-5.1 → MiniMax, so if one provider rate-limits, the agents degrade gracefully instead of going dark.

Workspace SeparationWorkspace Separation

Each agent has its own workspace on the gateway machine with its own SOUL.md (personality), MEMORY.md, and AGENTS.md. The workspace files get injected into every conversation — that's how personality persists across sessions.

The agents on ForgeMini have their workspaces on MacMax (the gateway reads them for context injection) but route their shell commands to ForgeMini. Took me a minute to figure that out — if you point the workspace path to the remote machine, the gateway can't read the files and the agent just... stops responding. Workspace = local to gateway, exec = remote node.

What Actually Works WellWhat Actually Works Well

Agent-to-agent communication. Naomi (PM) can assign tasks to Vandana (dev) by tagging her in Discord. Vandana does the work and reports back. I don't have to relay messages between them.

Personality specialization. Each agent's SOUL.md shapes how they approach problems. Vandana writes code differently than Lori reviews UI. Same underlying model, different perspective.

Cost predictability. Flat subscription means I never worry about a runaway agent burning through API credits. They can be as verbose as they need.

Mixed hardware. The Mac Mini cost $599 and draws 10 watts. It handles exec for 3 agents without breaking a sweat. You don't need expensive cloud VMs for everything.

What Bit MeWhat Bit Me

• allowBots: false is the default in Discord config. Bot accounts can't see messages from other bots unless you flip this. Spent an hour debugging why agents couldn't talk to each other.
• Discord channel permissions are separate from OpenClaw config. Even if OpenClaw allows an agent to respond in a channel, the bot's Discord role needs View Channel permissions. You'll hit this every time you add a new channel.
• Workspace paths must be local to the gateway. I tried pointing workspace paths to the remote Mac Mini paths. The gateway couldn't read SOUL.md, so the agent kept crashing with "stale-socket" restarts. Exec routing goes remote, workspace stays local.
• Per-agent model overrides don't inherit global fallbacks. If you set model.primary on an agent, you need to explicitly set model.fallbacks too. Otherwise the agent has zero fallback — one rate limit and it's dead.

The DashboardThe Dashboard

Built a monitoring dashboard that tracks all 4 nodes (CPU, RAM, disk), token usage per agent, and costs. It pulls metrics from local shell commands, SSH (for the Mac Mini), and prometheus exporters (for VPS nodes). The whole thing is a single HTML file + a Node.js server.

It shows subscriptions separately from hardware CapEx — because knowing your monthly burn rate and your one-time hardware investment are different questions.

Would I Recommend This?Would I Recommend This?

If you're running one Claw and it's working, don't overcomplicate it. But if you're finding yourself context-switching between "help me code this" and "review this design" and "what's the status on that project" — splitting into specialized agents is a real improvement. Each agent stays focused, context stays clean, and you stop re-explaining who you are every few messages.

The sweet spot is probably 3-4 agents for most people. I went to 8 because I wanted to see how far it scales. The answer: pretty far, but you'll spend a day debugging Discord permissions.

Running OpenClaw on a MacBook Pro + Mac Mini + 2 VPS nodes. Total cost: ~$318/month for 8 specialized agents.