In this codelab, you will install Docker, set up OpenClaw (an open-source, self-hosted AI assistant) in a container, and connect it to a local Ollama model (Gemma) so you can run AI-driven research workflows entirely on your own machine — no cloud API costs required.

Author: Geeta kakrani
GDE IN AI & TPU

What you'll learn

Prerequisites

Positive : Docker is optional for OpenClaw in general, but this codelab uses the containerized setup so you get an isolated, easy-to-clean-up environment.

  1. Open your browser and go to docker.com/products/docker-desktop
  2. Click Download for Mac – Apple Silicon (important: not the Intel version)
  3. Open the downloaded .dmg file and drag the Docker icon into your Applications folder
  4. Open Docker Desktop from Applications. Approve any system permission prompts with your Mac password
  5. Accept the terms of service. You can skip creating a Docker account for local use

Verify the install

Open Terminal and run:

docker --version
docker run hello-world

If you see Hello from Docker!, your installation is working.

Docker demo

Negative : If a popup says "Terminal would like to access data from other apps", click Allow — this is a normal macOS permission prompt for the Docker credential helper. 

mkdir -p ~/openclaw-project
cd ~/openclaw-project
git clone https://github.com/openclaw/openclaw.git
cd openclaw

Negative : If git isn't installed yet, macOS will prompt to install Command Line Developer Tools. Click Install, wait a few minutes, then re-run the clone command.

  1. Download Ollama from ollama.com/download if not already installed
  2. Pull the Gemma model you want to use:
ollama pull gemma3
  1. Verify it downloaded:
ollama list
  1. Confirm Ollama is reachable on your host:
curl http://127.0.0.1:11434

You should see Ollama is running.

Ollama is running

Positive : If ollama serve gives an "address already in use" error, that's fine — it means Ollama is already running in the background.

From inside the openclaw folder:

export OPENCLAW_IMAGE="ghcr.io/openclaw/openclaw:latest"
./scripts/docker/setup.sh

Walk through the interactive wizard:

Prompt

Selection

Personal-by-default disclaimer

Yes

Setup mode

QuickStart (recommended)

Model/auth provider

Skip for now (we'll configure Ollama separately)

Default model

Keep current

Select channel

None / Skip

Search provider

Skip for now

Configure skills now?

No

Enable hooks?

Skip for now

The script will build/pull the image and start the container.

Verify the container is running

docker compose ps

You should see openclaw-openclaw-gateway-1 with status Up ... (healthy).

OpenClaw installation OpenClaw status

Containers can't reach your Mac via 127.0.0.1 — that address means the container itself inside Docker. Use Docker's special hostname instead: 

http://host.docker.internal:11434

Run the interactive configure wizard to set this correctly:

docker compose run --rm -it openclaw-cli configure

In the wizard:

  1. Where will the Gateway run? → Local (this machine)
  2. What do you want to configure? → Model
  3. Search/select Ollama
  4. Ollama auth method → Ollama
  5. Ollama mode → Local only
  6. Ollama base URL → clear the default and type: 
    http://host.docker.internal:11434
  7. Pick your model (e.g. gemma3 or gemma4:e4b) from the picker
  8. Select Done to exit the wizard

Select model screenshot

Provide the required Ollama auth key

Ollama's local API needs a placeholder API key to be registered as a provider (any value works — it's not a real secret): 

echo 'OLLAMA_API_KEY=ollama' >> .env
docker compose up -d --force-recreate openclaw-gateway

Verify the variable made it into the container:

docker compose exec openclaw-gateway env | grep OLLAMA

Get your gateway token:

cat .env | grep TOKEN

Open in your browser, replacing the token placeholder with your actual value: 

http://127.0.0.1:18789/?token=YOUR_TOKEN_HERE

Negative : Keep your gateway token private. It grants full control of your OpenClaw instance. Never paste it into chat, screen-share, or share it with anyone — including AI assistants helping you debug.

If you see "Auth did not match"

This usually means a stale token. Fetch a fresh one:

docker compose run --rm openclaw-cli dashboard --no-open

If the token still doesn't work after recreating the container, check whether an old value is cached in your shell session: 

echo $OPENCLAW_GATEWAY_TOKEN
unset OPENCLAW_GATEWAY_TOKEN
docker compose up -d --force-recreate openclaw-gateway

In the Control UI chat, confirm the active model and provider:

/models

Switch to your local model if needed:

/model ollama/gemma3

Send a test message:

hi

You should get a response generated entirely by your local Gemma model — no external API calls, no cost.

Final output

Try a real research workflow

Find and summarize 3 recent approaches to AI-driven research workflows, with sources.

Observe how the agent breaks the task into steps, gathers information, and structures its output — this is the core capability worth demonstrating live.

Symptom

Likely cause

Fix

Ollama could not be reached at http://127.0.0.1:11434 (from inside setup wizard)

Wizard used container-local address

Manually enter http://host.docker.internal:11434

Unknown model: ollama/...

Missing Ollama API key

Add OLLAMA_API_KEY=ollama to .env, recreate container

Auth did not match in Control UI

Stale or mismatched token

Regenerate token via dashboard --no-open, force-recreate the gateway, clear any exported shell variable

Mac becomes slow/hangs

Disk nearly full from Docker images + Ollama models

Run docker system prune -a, remove unused Ollama models, restart the Mac to clear stale APFS snapshots

Cannot connect to the Docker daemon after a Mac restart

Docker Desktop didn't auto-start

Run open -a Docker and wait for it to fully load

You now have a fully self-hosted AI research assistant running locally — OpenClaw orchestrating a local Gemma model via Ollama, all inside an isolated Docker container with no cloud dependency or per-query cost.

What you learned

Next steps