Quickstart

Pull the image and start the container with two named volumes so your settings and chat history survive restarts. The two-e lines match the image defaults and are shown explicitly so you can see where state lives inside the container.

docker pull ghcr.io/alexk-dev/golemcore-bot:latest

docker run -d \
  --name golemcore-bot \
  --shm-size=256m \
  --cap-add=SYS_ADMIN \
  -e STORAGE_PATH=/app/workspace \
  -e TOOLS_WORKSPACE=/app/sandbox \
  -v golemcore-bot-data:/app/workspace \
  -v golemcore-bot-sandbox:/app/sandbox \
  -p 8080:8080 \
  --restart unless-stopped \
  ghcr.io/alexk-dev/golemcore-bot:latest

Unable to find image 'ghcr.io/alexk-dev/golemcore-bot:latest' locally
latest: Pulling from alexk-dev/golemcore-bot
Digest: sha256:7b2c9e5a1f...
Status: Downloaded newer image for ghcr.io/alexk-dev/golemcore-bot:latest
8f3e6a1c4d2b9e7fa5b3c1d0e8f7a6b5c4d3e2f1a0b9c8d7e6f5a4b3c2d1e0f9

On its very first boot the bot prints a one-time admin password to its own log. Grab it with:

docker logs golemcore-bot 2>&1 | grep "Initial admin password"

Initial admin password: Xf3kP9-qLmN2w7R

Open http://localhost:8080 in your browser. Sign in with the username admin and the password you just copied. You land on the onboarding screen.

Connecting a model is three short pages under Settings, in order. Each one does one thing: register the provider, tell the bot which model of that provider it is allowed to call, then assign that model to a tier.

1. Settings → LLM Providers. Click Add provider, pick your provider (OpenAI, Anthropic, Google, or any OpenAI-compatible endpoint), paste the API key, and save. This only registers the credentials — no model is available yet.
2. Settings → Model Catalog. Click Add model and pick the concrete model you want from the provider you just added — for example openai/gpt-5.1. The catalog is the list of models the bot is allowed to reach; the router can only point tiers at what lives here.
3. Settings → Model Router. Point balanced, smart, coding, and deep at the model you just added. For a first run it is fine to pin all four named tiers to the same model — you can split them later once you have stronger models for reasoning or code.

When to point each tier at a different model — and how the dynamic upgrade step decides to switch to coding mid-turn — is covered in Model Routing.

Open Chat. Type something simple — for example, Say hello in one sentence. The bot streams its reply in real time, with any tool calls and progress visible inline. The exact wording of the response depends on the model you wired up, but a short greeting confirms the full path (provider → catalog → router → chat) is working.

That is the full quickstart. The bot is running, the dashboard is yours, and your first conversation is in the books. Use Sessions and Logs if you want to look at what just happened in detail.