The Premise
I'm running a one-person hardware business. I design products, 3D print them, and sell them online. I don't have a marketing department. I don't have a customer support team. I don't have an ops person managing my email, calendar, and Shopify admin.
What I have is OpenClaw — an open-source, self-hosted gateway that connects AI models to my messaging apps, file system, browser, and the internet. It runs as a daemon on my MacBook Pro right now, with a dedicated Mac Mini on the roadmap for always-on operation. People are already running these setups on Mac Minis, VPS boxes, and even Raspberry Pis.
In just the first two weeks of daily use, the agent has helped build an entire e-commerce website, set up automated email follow-up pipelines, scaffolded a second site for a side project, written blog content, managed product photography, and even helped research dog breeders. Here's an honest breakdown of what it can actually do — department by department — and where the hype still outpaces reality.
What OpenClaw Actually Is
It's a Node.js process that sits between your chat apps and an AI model (Claude, GPT, Gemini — your pick). One Gateway handles:
It's open source (MIT), self-hosted, and your data stays on your machine. The only thing leaving is API calls to whichever model provider you choose.
Department by Department
What one AI agent can realistically handle for a small business.
Marketing & Content
Works todayThis is where the agent earns its keep. Content drafting is the single strongest use case for AI agents in a business context.
Writes full posts from a topic or outline. Quality is 80-90% there — you edit for voice and accuracy.
Drafts platform-specific posts (Twitter, Instagram, LinkedIn). Knows character limits and formatting.
Generates SEO-friendly copy from product specs and photos. Surprisingly good at this.
Maintains a markdown content plan. Suggests topics based on trends and your niche.
Scheduled cron job scrapes competitor sites weekly. Summarizes pricing, new products, positioning.
Web search for keyword opportunities, content gaps, backlink prospects.
Actual posting is a different story. You can use platform APIs (Twitter, Meta Graph) for automated posting, but browser-based posting to Instagram or LinkedIn is fragile. Best approach: agent drafts + schedules, you approve with a thumbs-up on WhatsApp, agent posts via API.
Customer Support
Needs effortWorks well for the 80% of support that's predictable. Falls apart on the 20% that requires judgment.
Agent knows your product inside-out from memory files. Answers sizing, compatibility, shipping questions instantly.
Queries Shopify API, gives customers real-time tracking info via WhatsApp/Telegram.
Gmail webhook → agent reads, categorizes, drafts responses. You review before sending.
Too nuanced. Agent can draft a response but shouldn't execute refunds autonomously. Human-in-the-loop required.
Hallucination risk. If the agent doesn't know, it needs to escalate, not guess.
Operations & Admin
Works todayThe unsexy work that eats your day. This is where automation saves the most time because no one wants to do it manually.
Cron fires at 8am: email summary, calendar for the day, overnight orders, any mentions or notifications. Delivered to WhatsApp.
Reads inbox, flags urgent items, drafts replies, archives noise. Gmail Pub/Sub webhook makes this near-realtime.
Checks upcoming events, sends reminders, can draft meeting agendas from context.
Queries Shopify for stock levels, alerts when something's running low.
Keeps workspace tidy. Organizes photos, documents, notes. Maintains its own memory system.
Product updates, price changes, collection management via API. Browser automation as fallback for things the API doesn't cover.
Research & Intelligence
Works todayAgents are research machines. This might be the most underrated capability.
Weekly cron job: scrape competitor sites, compare pricing, note new products, summarize changes. Delivered as a brief.
"What are people saying about X on Reddit?" — agent searches, reads threads, synthesizes findings.
"Find me 5 suppliers for X with pricing" — web search, fetch pages, compile comparison table.
"How does BLE HID work on nRF52840?" — deep dives into documentation, forums, GitHub repos.
Scheduled searches for industry keywords, new competitor products, regulatory changes.
Engineering & Development
Works todayThe agent can spawn coding sub-agents (Claude Code, Codex) for actual software development. This article's website was built by a sub-agent the main agent spawned.
"Build me a Next.js site with these pages" → spawns Claude Code → reports back when done.
Describe the issue, agent delegates to a coding sub-agent in a git worktree.
Point it at a PR, it reviews the diff and posts comments.
Check GitHub Actions, deploy status, error logs. Alerts on failures.
Generates and maintains docs from codebase. Keeps README and API docs current.
What Doesn't Work (Yet)
Not readyBeing honest about this matters more than selling the dream.
Agents make mistakes. Auto-processing refunds, modifying orders, or changing pricing without human approval is asking for trouble.
AI video (Runway, Sora, Pika) exists but quality is inconsistent. Not reliable enough for product videos you'd actually publish.
"Monitor inventory → reorder from supplier → update listings → notify customers" — each step works alone. Chaining them reliably breaks down.
The dream of an AI running your business while you sleep. Reality: agents hallucinate, misunderstand context, and make confident mistakes. Daily oversight is still required.
Works ~70-80% on well-structured pages. Falls apart with CAPTCHAs, dynamic loading, 2FA flows, and heavily JS-dependent admin panels.
The Hardware Setup
You don't need much. The agent runs as a lightweight Node.js daemon. The AI model runs in the cloud (Anthropic, OpenAI, etc.) — your machine just orchestrates. I'm currently running on a MacBook Pro, which works but sleeps when you close the lid. A dedicated Mac Mini is the move for always-on operation — low power, headless, and it just runs.
Also works on: any Linux box, a Raspberry Pi (slower), a VPS, or just your laptop (but it sleeps when you close the lid). The Mac advantage is iMessage integration — the agent can text you natively on your iPhone.
Monthly Running Costs
| Service | Cost | Notes |
|---|---|---|
| Anthropic Claude API | $20-100 | Scales with usage. Opus is pricier. |
| Brave Search API | Free-$5 | 2K queries/mo free tier |
| Shopify | $39 | API included with plan |
| Domain + hosting | $0-20 | Vercel free tier, $10/yr domain |
| Total | $60-165/mo | vs. hiring even one part-time employee |
Getting Started
Install & onboard
npm install -g openclaw
openclaw onboardWalks you through API key setup, channel config, and workspace creation.
Start the Gateway
openclaw gateway startRuns as a daemon. Auto-restarts on crash. Web UI at localhost:3007.
Connect a messaging app
# WhatsApp: scan QR in web UI
# Telegram: paste BotFather token
# Discord: add bot application tokenNow your agent lives in your pocket. Text it like a coworker.
Write SOUL.md
# In your agent's workspace
# Define personality, boundaries, capabilities
# This is the highest-leverage file you'll writeA good soul file transforms a generic chatbot into a collaborator that knows your business.
Set up cron jobs
# Morning briefing at 8am
# Weekly competitive research on Monday
# Daily email triage
# Calendar remindersThe agent becomes proactive. It checks in, does background work, and alerts you when something matters.
The Honest Take
An OpenClaw agent on a Mac Mini doesn't replace employees. It replaces the need for employees when you're a solo founder or tiny team. The difference matters.
The agent handles the 80% of work that's predictable: drafting content, triaging email, answering common questions, running scheduled checks, maintaining documentation, doing research. You handle the 20% that requires judgment: approving content before it posts, handling complex customer issues, making strategic decisions, quality-checking the agent's work.
The best mental model isn't "AI employee" — it's AI amplifier. You're still the brain. The agent is an extra pair of hands that never sleeps, never forgets, and can do ten things at once.
The businesses that will win with AI agents aren't the ones that automate everything. They're the ones that figure out which 80% to delegate and which 20% to protect.
Links: openclaw.ai · GitHub · Docs · Discord
This article was researched, drafted, formatted, and deployed by the same OpenClaw agent it describes. It also built the website you're reading it on, helped find a Rhodesian Ridgeback breeder earlier today, read a receipt screenshot to explain buyer protection fees, and scaffolded the entire site in under 10 minutes. We've been working together for two weeks.