Insights

69 AI Automations a Solo Business Runs on a Mac mini Server and Claude Code

6 min read#ai-automation#workflow-automation#claude-code#mac-mini-server#solo-business#discord-notifications

Who this is forSolo business owners and freelancers who want to cut repetitive work with AI but aren't sure what to automate first.

TL;DR: The MacBook isn’t always on, so I keep a Mac mini as a home server and run 69 automations built with Claude Code. Splitting each automation into five slots (trigger, Claude, memory, deploy, and arrival) shows where to start in your own work.

Contents

  1. The MacBook sleeps, so I got a Mac mini
  2. The first automations were built with n8n
  3. Reading automations as five slots
  4. Level 3: slots added one at a time
  5. Maintenance matters more than building
  6. A first automation to try tomorrow

I’m a solo business owner. As of October 20, 2026, 69 automations run on my two computers: 11 on the MacBook and 58 on the Mac mini. That number includes scheduled jobs and always-on services (including servers and backups), so not all of them call an AI.

This post shows how I build and run those 69. The main point isn’t any single automation, but how to read any of them with the same frame.

1. The MacBook sleeps, so I got a Mac mini

The MacBook isn’t always on. I carry it around, close the lid, and it goes to sleep. So I keep a Mac mini at home as an always-on home server. I build with Claude Code on the MacBook, and the Mac mini is where the things I build actually run.

Overall architecture diagram of building on the MacBook and running 24/7 on the Mac mini. I build with Claude Code on the MacBook and push the code to GitHub with git push. Eleven jobs also run on the MacBook. The Mac mini pulls the code from GitHub every hour, and 58 scheduled jobs and services respond to fixed times, messages, and emails. When judgment is needed, it asks claude -p once, and records are kept in SQLite. Results go out as Discord alerts and replies, a Google Sheets list, and a public address connected through Cloudflare.
GitHub connects the place where things are built (MacBook) with the place where they run (Mac mini). The Mac mini pulls the code every hour.

I don’t keep the 69 in my head. Every day, I scan both computers and collect everything into one Google Sheet.

Partial view of the automation list in Google Sheets. Columns include Machine, Label, Status, Schedule, and Last run. Each row shows a job running on the MacBook with an active status, a schedule such as daily at 07:00 or weekly on Sunday at 23:00, and its last run time. Two rows show exit 12 and exit 1 warnings in the status column, highlighted in red. One job name is blurred.
A row with the ⚠exit warning in the status column is a job whose last run failed.

If you’re new to servers, start with What is a server?. To connect to your home Mac mini from outside, see how to use Tailscale.

2. The first automations were built with n8n

I built my first automations with n8n. They covered household bookkeeping, government-run startup support programs in South Korea, and weekend festivals and events for parents with young kids.

One of them was a chatbot to help with grocery shopping. I built it in half a day by connecting farm produce price data to n8n’s chat node and AI agent node. It started from this thought:

“I’d be happy with just aggregate information on the items I’m interested in..”

(the author’s LinkedIn post, August 29, 2025)

The author's LinkedIn post from August 29, 2025. At the top is the title 'How I built a grocery-shopping AI agent with n8n in half a day' and the problem statement: after buying a zucchini for KRW 1,500 at a supermarket and seeing it sold for KRW 700 a few months later, the author felt the need for farm produce price data. The middle is trimmed. At the bottom is an n8n workflow screen with a chat trigger, AI Agent, OpenAI model, memory, HTTP request, and Google Sheets nodes connected in a line. The post shows 203 likes, 4 comments, 23 reshares, and 19,860 impressions.
A record from the first automation period. Below is the n8n screen I built then and the response this post received.

3. Reading automations as five slots

Automations look different from each other, but when you break them down, they split into five slots. I fill in these slots one at a time when I build something new.

Slot What it asks What I use in my automations
① Trigger When does it start? Fixed times, messages, emails, web form submissions
② Claude Who builds it and makes the judgment? Chatting with Claude Code while building, claude -p while running
③ Memory Where is it stored? Google Sheets, SQLite, Supabase
④ Deploy Is there an address others can reach? Cloudflare
⑤ Arrival Where do I see the results? Discord alerts, web page, sheet

③ Memory and ④ Deploy are added only when needed. That’s why the number of slots filled in tells you the stage an automation has reached.

4. Level 3: slots added one at a time

Three-level architecture diagram. Level 1, the notifier bot: at a fixed time, Claude makes a judgment and sends an alert to Discord. Level 2, the memory bot: memory is added, so Claude makes judgments while going back and forth with memory and sends an alert to Discord. Level 3, the service with its own address: visitors come in through a Cloudflare public address, and when needed, Claude processes the work while going back and forth with memory and shows the result on the same screen.
Memory (③) is added at Level 2, and the public address (④ Deploy) is added at Level 3. At Level 3, the trigger also becomes the visitor.

Level 1: Instructor mentoring job radar

This is a bot that finds teaching gigs. Every Monday at 7:50 a.m., it collects postings from two job sites and drops any it has already sent. For each new posting, claude -p assigns a star rating, and Discord cards arrive in order from highest to lowest rating. It keeps only enough records to avoid sending the same posting twice.

Instructor mentoring job radar card that arrived in Discord on Monday, October 19, 2026, at 7:51 a.m. It summarizes 5 new postings this week and 17 reviewed, with star rating criteria based on hourly pay, workload, location, and topic. The first posting's hourly pay, workload, location, role and headcount, course materials, deadline, and applicant count are listed in separate fields. The posting title and company are blurred.
The card that arrived on the morning of October 19. Five postings worth a fresh look, out of the 17 reviewed this week, came in star-rating order.

Level 2: Automatic expense sorting and a family Discord assistant

For expense sorting, I press the export button once in my household ledger app, and the Mac mini handles the rest. It sorts by store-name rules and shopping-mall order emails first. Only the expenses left over go to claude -p, which infers their purpose.

Automatic expense sorting diagram. When I press export once in the app, a ledger export email arrives. n8n on the Mac mini checks the email periodically and sorts first by store-name rules and order emails. Only what's left goes to claude -p, which infers the purpose. The classifications are saved to the expense database (SQLite), and then a Discord result card is sent.
Discord completion card for the household expense pipeline, received on September 7, 2026, at 10:27 p.m. Each step shows a completion mark and a count: receiving the export, loading, decision tagging, matching orders from Coupang (Korean online shopping platform), Kurly (Korean grocery delivery), and Naver Pay (Naver's payment service), applying purposes, LLM purpose inference, generation and verification, and deployment. The monthly variable amount and file path are blurred.
Each step sends a card showing how many items it handled.

The family Discord assistant works like this: when family members post a message in Discord, a program that stays running on the Mac mini picks it up and writes a reply with claude -p. Conversations are stored in SQLite, and past chats can be searched when needed.

Level 3: Writing feedback site and homepage inquiry form

The writing feedback site is one where a single user comes in from a phone. It runs on the Mac mini, but I connected a public address through Cloudflare. The moment someone else visits, the ④ Deploy slot turns on.

Writing feedback site diagram. A user enters from a phone through a Cloudflare public address to the web app on the Mac mini. The web app checks a PIN, receives the draft, saves a work record in SQLite, and claude -p performs the feedback and record cleanup, then shows the results on the same screen.

The homepage inquiry form is the case where the memory slot lives outside the Mac mini. When a visitor submits an inquiry, a row is added to an inquiry table in Supabase, and after saving, n8n sends me a notification. Claude isn’t involved in this flow, and the result comes to me as a notification rather than on the same screen.

5. Maintenance matters more than building

I don’t build everything from scratch, either. For things Notion or Google Sheets can already handle, I keep using them. I split the rest between work a home server is enough for and work that needs a commercial service like Supabase.

Memory slot Nature Examples from my automations
Google Sheets A list people look at and edit by hand Work waiting list, automation list
SQLite A file-based database inside one computer Expense database, assistant’s conversation memory
Supabase A database on the internet Homepage inquiry form

I move things to Supabase when several people write at the same time, when login is required, or when a web page without a server needs to read the data directly.

I’ve also retired some automations.

  • Photo migration bot: After moving 43,615 OneDrive photos to Google Photos, it still had no stop condition. It kept retrying the remaining files every 5 minutes and ran for 39 hours without progress before it stopped.
  • Multiple weekly cards: Retrospective cards that used to arrive separately on Monday morning and Sunday night were merged into a single Monday weekly report card.

6. A first automation to try tomorrow

69 isn’t a small number, but it will keep growing. If you’re starting out, I suggest beginning with one Level 1 automation: a bot that collects web data with Claude Code and sends the results to you on Discord.

BuildnWrite helps teams build AI agents that keep running. About BuildnWrite ›