DEV Community

linou518
linou518

Posted on

Building a Dedicated AI Assistant Environment with a Single Mini PC

#homelab #ai #selfhosted #docker

Building a Dedicated AI Assistant Environment with a Single Mini PC

When starting a new company, I hit a contradiction: "Don't want to install random software on the work laptop" versus "Want to use an AI assistant." The fix: get one Mini PC and put everything on it.

Background

Company security policy prohibits installing extra software on the work Mac. But I needed an AI assistant (OpenClaw) for daily work. Since I work from home and everything stays on the local network, I decided to add a Mini PC (GMK) as a dedicated environment.

What I needed:

  • File sharing (exchange documents with the work Mac)
  • Chat UI (an interface to talk to the AI assistant)
  • The AI agent itself

Architecture 

Work Mac (Finder / Browser)
    │
    ├── smb://192.168.x.x/share  → Samba (file sharing)
    └── http://192.168.x.x:8065  → Mattermost (chat UI)
                                        ↕
                                   OpenClaw Agent ("Joy")
Enter fullscreen mode Exit fullscreen mode

Everything runs on the Mini PC (Ubuntu 24.04, 16GB RAM). LLM inference happens in the cloud (Anthropic API), so local specs can be minimal.

Step 1: Samba — File Sharing

A simple file share accessible from Mac's Finder via Cmd+K → smb://192.168.x.x/share:

sudo apt-get install samba
sudo mkdir -p /home/openclaw/mdk-share
Enter fullscreen mode Exit fullscreen mode

Add a share section to /etc/samba/smb.conf:

[mdk]
   path = /home/openclaw/mdk-share
   browseable = yes
   read only = no
   valid users = openclaw
Enter fullscreen mode Exit fullscreen mode 
sudo smbpasswd -a openclaw
sudo systemctl restart smbd
Enter fullscreen mode Exit fullscreen mode

That's it. Drag and drop files from the Mac.

Step 2: Mattermost — Chat UI (Docker)

Mattermost serves as the chat interface to the AI assistant. It's a Slack-like tool that you can self-host — that's the key.

# docker-compose.yml
services:
  db:
    image: postgres:15-alpine
    environment:
      POSTGRES_DB: mattermost
      POSTGRES_USER: mmuser
      POSTGRES_PASSWORD: <your-password>
    volumes:
      - pgdata:/var/lib/postgresql/data

  mattermost:
    image: mattermost/mattermost-team-edition:latest
    depends_on:
      - db
    environment:
      MM_SQLSETTINGS_DRIVERNAME: postgres
      MM_SQLSETTINGS_DATASOURCE: "postgres://mmuser:<password>@db:5432/mattermost?sslmode=disable"
      MM_SERVICESETTINGS_SITEURL: "http://192.168.x.x:8065"
    ports:
      - "8065:8065"
    volumes:
      - mmdata:/mattermost/data

volumes:
  pgdata:
  mmdata:
Enter fullscreen mode Exit fullscreen mode 
docker compose up -d
Enter fullscreen mode Exit fullscreen mode

Open http://192.168.x.x:8065 in a browser, complete the initial setup, create an admin account, then create a Bot Account (@assistant) and save the token.

Step 3: OpenClaw — The AI Agent

Install OpenClaw and connect it to the Mattermost bot:

npm install -g openclaw
openclaw init
openclaw plugins install @openclaw/mattermost
Enter fullscreen mode Exit fullscreen mode

Define the Mattermost connection in openclaw.yaml:

channels:
  mattermost:
    baseUrl: http://localhost:8065    # ← "baseUrl", NOT "url"!
    botToken: <bot-token>
    dmPolicy: open
Enter fullscreen mode Exit fullscreen mode

Gotcha: The config key is baseUrl, not url. The docs and actual behavior don't quite match — cost me 30 minutes.

Write the agent's identity file (SOUL.md) and start with openclaw gateway start. When "connected as @assistant ✅" appears in Mattermost, you're done.

Step 4: Giving the Agent an Identity

OpenClaw agents define "who they are" through a SOUL.md file. I set this one up as a company tech assistant called "Joy":

  • Expertise: Cloud architecture (AWS/GCP/Azure), data engineering, IaC
  • Personality: Professional, results-driven, concise
  • Model: Claude Sonnet

Type @joy hello in Mattermost and it responds in character.

Time and Cost

  • Build time: About 1.5 hours (Samba 10 min + Mattermost 30 min + OpenClaw 30 min + debugging 20 min)
  • Hardware: GMK Mini PC (~$200, repurposed from existing inventory)
  • Running cost: LLM API fees only (electricity is a few dollars/month)

Lessons Learned

  1. UI choice matters. Only developers can work in a terminal. Adding Mattermost lets non-engineers use the AI assistant through chat.
  2. Self-hosting feels secure. Company data doesn't flow to SaaS. The LLM API calls do go out, but chat history and files stay local.
  3. A Mini PC is enough. LLM inference is in the cloud, so local hardware just relays API calls and serves the chat UI. 16GB RAM is plenty.
  4. baseUrl is not url. Trust the docs, but read the source when things don't work.

Conclusion

The need was "use an AI assistant without polluting the work laptop." The solution: Mini PC + Samba + Mattermost + OpenClaw. Built in 1.5 hours, near-zero additional cost (repurposed hardware).

This setup works because it's remote work on a local network. Down the road, VPN + reverse proxy could enable remote access.

Top comments (0)