How to Self-Host AI as Workspace with Docker - Step-by-Step Tutorial

Self-Host AI as Workspace with Docker — A Step-by-Step Tutorial

I remember the first time I tried setting up a self-hosted AI chat client. Three hours, six failed docker-compose attempts, and a lot of swearing later, I had something that sort of worked. AI as Workspace was the complete opposite experience: I was up and running in under 3 minutes. Let me show you how I did it — including the mistake that cost me an extra 10 minutes so you don't make it.

Prerequisites (What You'll Need)

• Docker + Docker Compose installed (I used Docker 24 on Ubuntu 24.04, but any recent version works)
• ~1GB RAM available (idle is ~120MB, but with a loaded model it can spike)
• An API key from OpenAI, Anthropic, Google, or any OpenAI-compatible provider
• ~5 minutes of your time — I promise it won't take longer

Step 1: Pull the Docker Image

I like to pull the image first so I can see the download progress and verify everything's fine before running:

docker pull krytro/aiaw:latest

This took about 30 seconds on my connection — the image is only ~80MB compressed, which is refreshingly small compared to other AI tools (I'm looking at you, ComfyUI's 8GB).

Step 2: Run the Container

Here's where I made my mistake. The first time, I ran it without mounting a volume, thinking "it's just a chat app, what data could there be?" Turns out — everything. Your workspaces, plugin configs, conversation history, Artifact content — all of that lives in /data inside the container. Without a volume mount, you lose it all when the container restarts. Don't be me.

docker run -d \
  --name aiaw \
  -p 8080:8080 \
  -v ./aiaw-data:/data \
  krytro/aiaw:latest

After running this, check the logs to confirm it started:

docker logs aiaw --tail 20

If you see something like "Server listening on http://0.0.0.0:8080", you're golden. If not, give it another 10 seconds — the first startup takes slightly longer as it initializes the database.

Step 3: Open the App and Configure Your First Workspace

Open your browser to http://localhost:8080. You'll be greeted by a clean setup screen asking for your API key and preferences. Here's what I did:

1. Add your API key — I use OpenAI's key format (sk-...), but it accepts Anthropic and Google keys too. The MCP integration even lets you use local models through Ollama.
2. Create your first workspace — I called mine "General" and set GPT-4o as the model. The workspace naming is important: you'll thank yourself later when you have 5 workspaces.
3. Browse the plugin marketplace — Click the puzzle icon in the sidebar. I installed the "Web Search" plugin (gives the AI browsing ability) and the "Code Interpreter" plugin. Both installed in under 5 seconds.

Step 4: Quick Test — Ask It Something Real

Don't waste time on "hello world." Ask it something concrete. I tested with: "Create a comparison table of the top 5 open-source vector databases in 2026, including stars, license, and Docker image size."

The response came back formatted as a markdown table in the chat, plus a persistent artifact I could reference later. The artifact feature is something I didn't appreciate until I used it — now I can't live without it.

Step 5 (Optional): Docker Compose for Persistence

If you want a proper setup that survives reboots and is easier to manage, here's the docker-compose.yml I use:

version: '3.8'
services:
  ai-as-workspace:
    image: krytro/aiaw:latest
    restart: unless-stopped
    ports:
      - "8080:8080"
    volumes:
      - ./data/ai-as-workspace:/data

Save this as docker-compose.yml and run docker compose up -d. The restart: unless-stopped means it auto-starts after a server reboot — which is great if you're running this on a home server or VPS.

Common Pitfalls (And How I Fixed Them)

Final Thoughts

Self-hosting AI as Workspace was genuinely one of the smoothest Docker experiences I've had with an AI tool. The image is tiny, the setup is straightforward, and the multi-workspace concept actually delivers on its promise. If you've been putting off moving away from ChatGPT's single-feed interface, this is your weekend project — and it'll take you less time than a coffee break.