Building an AI Personal Server
with a Used Intel Mac mini for 400,000 Won

A practical account of native Ubuntu Linux installation + Claude Code + OpenClaw operation

March 2026 | Kim Kyung-jin

There is an Intel Mac mini sitting in one corner of my study. A 2018 model, i5 processor, 32 gigabytes of RAM. After Apple announced the M1 chip, this machine slipped out of the world's attention. The used price is around 400,000 won. Someone else might put it in a drawer or list it on Karrot Market, but I installed Ubuntu Linux on it and turned it into an AI server. Now this machine runs 24 hours a day as my personal AI assistant.

After trying several setups, the best combination I have found so far is this. Put a 1TB SSD in an Intel Mac mini, install Ubuntu Linux natively, and run Claude Code and OpenClaw on top of it. The hardware price is the lowest, but cloud-based AI tools run on it the cleanest and fastest.

Why Linux? The moment of rediscovery

The Intel Mac mini with macOS installed was honestly slow. The operating system itself had become heavy. Apple released a new macOS every year and optimized it for the latest chips, while Intel Macs became more and more burdened. But once Linux was installed, the situation changed. In a server environment running only a text terminal without a GUI, this machine felt snappy. The room of 32 gigabytes of RAM came alive, and the 1TB SSD opened fully for work.

AI tools such as Claude Code and OpenClaw do not perform heavy computation locally. They send commands to large language models, LLMs, in the cloud and receive results. Heavy calculation is done by Anthropic or OpenAI servers, and the Intel Mac mini only serves as a terminal that relays those commands. That is why the old Intel i5 processor is not a bottleneck. A GPU is not needed either. Network connectivity and stable power are enough.

The copy-paste problem: the Mac terminal was the answer

Linux servers have one inconvenience: copy and paste. On a GUI-free text terminal screen, you cannot use the mouse, and copying and pasting long authentication URLs or token strings is impossible. I hit this problem during the OAuth authentication step while installing OpenClaw. Typing a long authentication key without a single wrong character was not realistic.

The solution came from the MacBook in the same house. Open the Mac terminal and connect to the Intel Mac mini through SSH, secure remote access, and you can operate the Linux server directly from the MacBook screen. Since you can use the MacBook keyboard and mouse, copy and paste is free. The command is this one line.

ssh userID@192.168.219.110

Because both machines are connected to the same home router, you can connect immediately without separate security settings. From then on, I performed all installation and setup work from the MacBook terminal. It is a way to use the power of the Linux server while taking the comfort of operation from the Mac. In the end, if the MacBook is used only as a terminal remote control, you do not need a high-end machine.

Claude Code: a coding agent unfolding in the terminal

Claude Code is a terminal-based AI coding tool made by Anthropic. You do not need to open a browser. In the terminal, type claude to start it. It reads files, writes code, runs commands, and handles edits across multiple files at once.

Installation takes two minutes. One command is enough. But you are not plugging a keyboard directly into the Intel Mac mini to type this command. As explained above, you enter the command below while connected by SSH from the MacBook terminal to Ubuntu on the Intel Mac mini. What you see on the MacBook screen is the MacBook's terminal window, but the command running inside it runs on Ubuntu Linux on the Intel Mac mini. The same applies to all installation commands later in this article. The MacBook is the remote control, and the real work happens on the Intel Mac mini.

curl -fsSL https://claude.ai/install.sh | bash

You do not need to install Node.js first. Since 2026, a native installation method has been introduced, and dependency problems have disappeared. A Claude Pro subscription, $20 per month or higher, is needed, but one subscription lets you use both Claude on the web interface and Claude Code in the terminal, so the cost efficiency is good. To add an honest confession, you can start with the $20 plan. But once you begin using Claude Code, you lose track of day and night. "I will just do this and sleep" becomes 3 a.m. You ask it to organize files, then end up writing an automation script. You write the script, then a whole project appears. You run the project, then you hit the usage limit that resets every five hours. Then in the end you move up to Claude Max, $100 or $200 per month. I did too. I started at $20, and now I use the $200 plan. Whether that cost feels wasteful is something you will decide for yourself after trying it once.

Claude Code needs one more thing to work properly. Reference materials have to be inside Ubuntu Linux on the Intel Mac mini. Claude Code can read only documents inside the file system it can access. So files you usually worked on from the Mac have to be copied into the Intel Mac mini.

The copy method is simple. Use the scp, secure copy, command from the MacBook terminal. For example, to send one PDF file on the MacBook desktop to the Intel Mac mini, type this.

scp ~/Desktop/report.pdf kimkj@192.168.219.110:/home/kimkj/data/

If you want to send an entire folder, add the -r option.

scp -r ~/Documents/AI_research_materials/ kimkj@192.168.219.110:/home/kimkj/data/

Files in iCloud work the same way. On a Mac, the actual path of iCloud Drive is long, but you only need to know it once.

scp -r ~/Library/Mobile\ Documents/com~apple~CloudDocs/work_folder/ kimkj@192.168.219.110:/home/kimkj/data/

All these commands are entered from the MacBook terminal. You are transferring files from the MacBook to Ubuntu Linux on the Intel Mac mini over the network. Because both machines are connected to the same home router, transfer speed is fast. Even thousands of files can be moved in a few dozen minutes.

Once you put materials there, Claude Code can use them freely. Connect to the Intel Mac mini by SSH from the MacBook, run Claude Code, and speak to it in natural language. Let me show a few real scenes of use.

First, contact and phone-number search. I put an Excel file containing acquaintances' contact information inside the Intel Mac mini. If I say, "Find Attorney Kim OO's phone number," Claude Code opens the Excel file, searches that name, and tells me the number. Even in business-card data with thousands of rows, the result appears within seconds. There is no need to dig through a paper business-card binder, and I can also find people who are not saved in my smartphone contacts.

Second, context search through past materials. This is the most useful function. If I say, "I wrote something about the EU AI Act in 2024. Find the part where I mentioned the fine provisions," Claude Code searches thousands of files stored inside Ubuntu Linux on the Intel Mac mini and finds the relevant document. It finds keywords with a search command called grep, reads the file content, and understands the context. It tells me exactly which file and which paragraph contains the relevant content. It is content inside my personal materials that Google search could never find.

Third, creating connections between materials. If you know the knowledge management tool Obsidian, you will understand the idea of creating a web of knowledge by linking notes to one another. Claude Code plays that role. If I say, "Find my other documents related to this lecture material and organize how they are connected," it reads multiple files, analyzes intersections in the content, and explains the relationships. It can find the link between a column written three years ago and lecture notes made last month. In practice, you can copy all Obsidian markdown files into the Intel Mac mini and ask Claude Code to automatically create backlinks between notes.

Fourth, reading and summarizing PDF documents. Put in a report PDF hundreds of pages long and say, "Summarize the key content of Chapter 3," and it opens the file, reads it, and summarizes it. You can also ask it to write a Python script that selects only documents containing a specific keyword from 100 PDFs in a folder.

I currently keep more than 5,000 material files inside this Intel Mac mini and let Claude Code handle search, classification, and connection. All this work runs comfortably on an Intel i5. That is because the core of what Claude Code does is communicate with Anthropic's cloud AI model, while Ubuntu Linux on the Intel Mac mini only serves as a path for storing files and exchanging commands.

Uses keep expanding beyond this. When writing a book, you can put chapter drafts in folders and say, "Find where the arguments in Chapters 3 and 5 conflict," and it reads the two files and pinpoints where the content diverges. If you ask it to build the table of contents structure for a 30-chapter manuscript, it analyzes the flow of existing drafts and makes suggestions. Receipt organization works too. Put 250 receipt photos or PDFs in a folder and say, "Classify these by date and item and organize them in Excel," and it writes a Python script, reads file names and content, classifies them into subfolders, extracts amounts and dates, and exports a CSV file. Coding works the same way. If you say, "Make a Python script that crawls AI news every morning at 8 and sends it to Telegram," it writes the script, tests execution, and even registers it in cron, Linux's scheduled execution system. When making lecture materials, you can ask it to extract key content from existing PDF handouts and convert it into markdown slides, or put in a transcript of a recording and say, "Organize this as meeting minutes separated by speaker."

OpenClaw: a messenger becomes an AI control room

OpenClaw is an open-source AI agent that spread explosively among developers around the world in early 2026. It surpassed 240,000 GitHub stars and drew enough attention that NVIDIA CEO Jensen Huang mentioned it publicly. The core idea of this tool is plain. It uses messenger apps such as Telegram, WhatsApp, and Discord as the control interface for AI.

If you install OpenClaw on Ubuntu on the Intel Mac mini, this machine becomes a 24-hour AI gateway. I send commands to this server through Telegram. When I am out and send a smartphone message saying, "Summarize today's AI news," OpenClaw on the Intel Mac mini at home receives the command, executes it, and sends the result back through Telegram. I do not have to be sitting in front of a computer.

For OpenClaw's brain, you can connect models like ChatGPT through OAuth authentication. The setup I use is OpenClaw + ChatGPT 5.4. In Terminal, start initial setup with openclaw onboard and you will see step-by-step guidance. If you proceed while connected through SSH as described earlier, you can freely copy and paste the authentication URL.

A used MacBook is enough: the magic of SSH

When you connect to Ubuntu on the Intel Mac mini through SSH and run Claude Code, the screen is a 100% identical interface to installing and running Claude Code directly on a MacBook. The speed at which text appears, the way commands are entered, the process of reading and writing files, everything is the same. The only difference is that actual computation and file storage happen not on the MacBook but on Ubuntu Linux on the Intel Mac mini.

The implication is clear. The MacBook used for SSH access does not need to be an expensive latest model. A used MacBook Air, even a 10-year-old old MacBook, is fine. It only has to run one terminal app. Heavy work is handled by Ubuntu on the Intel Mac mini, and the MacBook merely serves as monitor and keyboard for that screen. Even without an M5 MacBook Pro, a used MacBook costing 200,000 won can provide the exact same experience.

Telegram connection: commanding AI from a smartphone

In March 2026, Anthropic officially introduced a feature called Channels to Claude Code. It lets Telegram, Discord, and iMessage be used as interfaces for Claude Code. It is the same concept as OpenClaw's Telegram connection, but it differs in that it is an official Anthropic plugin. I will show the setup step by step.

Step 1. Install Bun. Claude Code's Channels plugin runs on a JavaScript runtime called Bun. While connected by SSH to Ubuntu on the Intel Mac mini, enter the command below.

curl -fsSL https://bun.sh/install | bash

Step 2. Create a Telegram bot. Open the Telegram app on your smartphone and send a message to @BotFather. Type /newbot, and it asks for the bot name and username. The name is up to you, for example, "My AI Assistant," and the username must end with bot, for example, kimkj_ai_bot. When finished, BotFather gives you a token. It looks like 123456789:AAHfiqksKZ8... Copy it.

Step 3. Install the plugin in Claude Code. While connected to the Intel Mac mini through SSH, run Claude Code and enter the commands below in order.

/plugin marketplace add anthropics/claude-plugins-official
/plugin install telegram@claude-plugins-official
/reload-plugins

Step 4. Register the bot token. Inside Claude Code, enter the command below and paste the token received from BotFather.

/telegram:configure

Step 5. Relaunch in channel mode. Exit Claude Code and start it again with the command below.

claude --channels plugin:telegram@claude-plugins-official

Step 6. Pairing. On your smartphone, send any message to your bot in Telegram. The bot replies with a six-digit pairing code. Enter this code in the Claude Code terminal, and the connection is complete. From then on, every message sent through Telegram is delivered to Claude Code on the Intel Mac mini, and the response returns through Telegram.

It is good to use tmux so the session does not disconnect. If you start it with the command below after SSH access, Claude Code keeps running even if you close the terminal.

tmux new-session -d -s claude "claude --channels plugin:telegram@claude-plugins-official"

Auto mode and skipping permissions: fully autonomous execution

If you send a command through Telegram and Claude Code asks, "May I modify this file?" that is a problem. On a smartphone, there is no way to press the approval button. You have to go directly to the terminal and press yes. If you are out, that is useless.

There is a command that solves this problem.

claude --channels plugin:telegram@claude-plugins-official --dangerously-skip-permissions

The --dangerously-skip-permissions flag is exactly what its name says: an option that dangerously skips permission checks. Turn this on, and Claude Code immediately creates, edits, deletes files, and runs commands without asking for permission each time. If you send "run the test script" from Telegram, it runs immediately without an approval request and sends the result to Telegram.

There is a reason the word dangerously is in the name. You are handing AI full authority over the computer. It is not an option to use carelessly on a computer containing personal data. But the Intel Mac mini Ubuntu server is a machine set up only for AI work. It has no personal mail and no banking apps. In this environment, the risk of using this option is limited.

Claude Code also has auto-accept mode. With Claude Code running in the terminal, press Shift+Tab to switch between Plan Mode and Auto-accept Mode. In auto-accept mode, Claude Code judges on its own, reads files, writes code, and runs commands. It handles complex multi-step work continuously without human approval each time. It can write a Python script, execute it, automatically fix errors if they occur, and run it again, all by itself.

An honest conclusion: do you really need OpenClaw?

If you have read this far, one question may come up. If Claude Code connects to Telegram, supports auto mode, searches files, and runs code, is there any reason to install OpenClaw separately?

In practice, if you run Claude Code with a flat-rate account such as Claude Pro, $20 per month, it can replace almost everything OpenClaw does. Sending remote commands through Telegram, managing files, searching the web, writing scripts for automation. Claude Code alone can do it all. OpenClaw has the advantage that you can freely switch among multiple AI models and use more than 100 community skills, but the Claude Code + Telegram Channels combination is so complete that most individual users have almost no need to install OpenClaw separately.

The setup is this. Install Ubuntu on the Intel Mac mini, install Claude Code, and connect the Telegram channel. That alone completes a 24-hour personal AI assistant. There is no need to add OpenClaw; one Claude Pro subscription is enough.

File freedom: putting materials in with macFUSE and rsync

For AI to do its job, reference materials must be inside Ubuntu Linux on the Intel Mac mini. I copied the Mac's iCloud and local files into the Intel Mac mini over the home network. There are two methods.

First, use macFUSE + SSHFS. If you install macFUSE on a Mac, you can mount a Linux directory in the Mac Finder like a drive. You only have to drag and drop files as if an external hard drive were connected. In Finder, this drive appears with the name "macFUSE Volume 0 (sshfs)."

When macFUSE is first installed, it does not work immediately. That is because Mac security settings block external kernel extensions. Once you unlock this setting, you can use it freely afterward. The procedure differs by Mac type.

This applies to Intel Macs with a T2 security chip, such as Intel MacBook Pro and Mac mini models after 2018. When you run macFUSE for the first time after installation, a notification appears saying "System software was blocked." Go to System Settings -> Privacy & Security, and at the bottom you will see a message that software from developer "Benjamin Fleischer" was blocked. Press the Allow button and enter the administrator password. Restart the Mac, and macFUSE will work.

For Apple silicon Macs (M1 through M5), there is one more step. Apple silicon Macs block third-party kernel extensions by default, so you must lower the security level one step in Startup Security Utility. When you install macFUSE and run it for the first time, a popup appears saying "System Extension Blocked." In System Settings -> Privacy & Security, press the button to enable system extensions. After you enter your password, the Mac shuts down. Once the Mac is completely off, press and hold the power button (or Touch ID button) to enter recovery mode. From the menu at the top of the screen, choose Utilities -> Startup Security Utility. Select the startup disk, press the Security Policy button, and choose Reduced Security. Under that, check the box for "Allow user management of kernel extensions from identified developers." Enter your password and restart. After restarting, go back to System Settings -> Privacy & Security and press the button to allow software from "Benjamin Fleischer." Restart one more time and macFUSE will work. It looks complicated, but you only have to do it once. After that, when macFUSE updates, you only need to press the allow button.

Second, you can create an rsync script for automatic synchronization. I run a script through launchd, the Mac's scheduled execution system, every evening at 9 p.m. It automatically copies files modified within the past week from the Mac's Documents, Downloads, Desktop, and iCloud folders to the Intel Mac mini. If you set up SSH key authentication in advance, it runs unattended without password entry.

Claude Code can freely search and use the materials collected this way. If you say, "Find the AI regulation document I wrote last week," it searches files on the server and shows the relevant documents. This is the advantage of a local server that cloud AI alone cannot provide.

Cost calculation: this is the key

Here is the summary.

• Hardware: used Intel Mac mini, about 400,000 won (one-time investment)

• Operating system: Ubuntu Linux free

• Claude Code: start with Claude Pro at $20 per month; if you use it seriously, Max costs $100-200 per month

• OpenClaw: the software itself is free, with API costs of $6-30 per month depending on usage

• Electricity: about 3,000-5,000 won

If you buy an M4 Pro MacBook Pro and run a local AI model, it costs at least 3 million won. But with a 400,000 won Intel Mac mini plus a subscription fee of 30,000-50,000 won per month, you can use the latest large cloud models as they are. No matter how good a local model is, it cannot catch up with the latest cloud models. This is a structure where you spend less money and get higher performance.

One limitation and a security note

Input on the Intel Mac mini is quite inconvenient. Korean input is awkward in the Ubuntu text terminal itself, and there is no mouse. So you must work through SSH from the Mac terminal environment. The Mac terminal is, in practice, a remote control for the Linux server.

From a security standpoint, OpenClaw requires caution. This tool has permission to run shell commands, access files, and make network requests, so it can be dangerous if configured incorrectly. As of February 2026, several OpenClaw-related security vulnerabilities (CVEs) had been reported, to the point that NVIDIA released a dedicated security add-on called NemoClaw. I created a dedicated Linux user account to limit permissions and set a daily usage cap on the API key. If you put basic security measures in place, running it inside a home network is safe.

Who I recommend this combination for

If you have an unused Intel Mac mini at home, do not leave it in a drawer. Installing Ubuntu takes one USB drive and 30 minutes. Installing Claude Code on it takes 2 minutes, and installing OpenClaw takes 10 minutes. In one hour, you have an AI server running 24 hours a day.

I recommend this combination for people who use AI a lot but feel burdened by buying the latest equipment, people who want to assign work to AI through Telegram while outside, and people who want to store their own materials locally and have AI search them. Even without an expensive GPU server or the latest Mac, you can put the power of cloud AI in your own hands. The key is not whether the equipment is new, but where and how you place it.

Bonus: Tailscale VPN and the Great Firewall

If you install a VPN called Tailscale on this Intel Mac mini running Ubuntu Linux, you can safely connect to this home server even from overseas. Tailscale is a mesh VPN based on the WireGuard protocol. It is very easy to install, and the free plan can connect up to 20 devices. The command to install it on Ubuntu on the Intel Mac mini is one line.

curl -fsSL https://tailscale.com/install.sh | sh
sudo tailscale up --advertise-exit-node

The key is --advertise-exit-node on the second line. When this option is enabled, the Intel Mac mini becomes an exit node. If you turn on the Tailscale app on a smartphone or laptop and select this exit node, all internet traffic goes out through the Intel Mac mini at your home in Korea. From the outside, it is the same as connecting from Korea.

This is a major help for people studying or working in China. China's Great Firewall blocks access to foreign services such as Google, YouTube, Claude, ChatGPT, Telegram, and WhatsApp. Commercial VPNs are often blocked and unstable. But if an Intel Mac mini Ubuntu server is running at your home in Korea and Tailscale exit node is set up on it, you can use Korean internet as-is from China simply by turning on the Tailscale app on your smartphone.

A child studying abroad in Shanghai needs to access assignment files on Google Drive. They open the Tailscale app and select the home exit node in Korea. Google Drive opens. A KakaoTalk video call with parents does not cut out either. An office worker on a business trip in Beijing needs to ask Claude a work-related question. They turn on Tailscale. claude.ai connects. Sending commands through Telegram to Claude Code running on the Intel Mac mini at home also works. That is because Tailscale ties both sides into one private network.

Because Tailscale is based on WireGuard, it does not have traffic obfuscation, and it may be detected by the Great Firewall's deep packet inspection (DPI). Depending on region and timing, there have also been reports of Tailscale access itself being blocked. It is not a perfect solution. Still, many users report that Tailscale is more stable than ordinary VPNs, and stability improves further if you build your own DERP relay server. A realistic method is to use it alongside a commercial VPN or keep it as a backup for times when other connections fail.

If you install Ubuntu on a used Intel Mac mini bought for 400,000 won, add Claude Code, connect a Telegram channel, and install Tailscale, one machine runs 24 hours a day as an AI work server, a VPN exit for overseas access, and file storage. I confirmed directly that an old Mac mini sleeping in a drawer can do this much.