Quick start

The fastest path from a fresh install to an AI agent building and operating apps on your Atelier instance. Two routes: the bundled agent (zero setup) or your own agent (Claude Code, Cursor, a chat assistant). Either way, the agent drives Atelier through the same public interface.

1. Install Atelier

Follow the install guide. You end up with a running platform at a URL like http://atelier.local (or the domain you chose).

2a. The fast path: the bundled agent

If you just want to go, open the portal and choose Build with AI. Atelier runs a build agent (Claude Code or opencode) inside your own cluster with the skills already installed. Describe the app you want; it builds and ships it. Nothing to download. (An admin enables it once under Settings -> AI -> Bundled agent.)

2b. Bring your own agent

Prefer to drive Atelier from an agent you already run? Three steps, and step 2 is the one we actively want you to do.

Step 1 — mint a scoped token

In Atelier: Settings -> API tokens. Create a token with the role you want the agent to have (Developer to build and deploy apps, Viewer for read-only). It’s shown once; copy it.

Atelier teaches any skill-aware agent how to drive it through two Agent Skills:

atelier-build — author code + a Dockerfile and deploy it; Atelier builds the image and owns the Kubernetes plumbing.
atelier-operate — list, inspect, debug, and manage apps already running (logs, metrics, secrets, pause/resume/redeploy).

They are plain Markdown instructions — nothing executes — but they are exactly what your agent will be told to do: which API calls to make, where it pushes git, how it uses your token. So look before you install. Two ways:

Let your agent fetch and vet them for you (recommended). Paste this prompt to your agent — it downloads the skills, reads them in full, summarises what they do, and installs them only once you approve:

Download the Atelier agent skills from https://tryatelier.blob.core.windows.net/tryatelier/latest/skills.tar.gz and extract them somewhere temporary. Before doing anything else, read both SKILL.md files (atelier-build and atelier-operate) and the bring-your-own-code section in full, then summarise for me in plain English: what they instruct you to do, which network calls and git operations they involve, how they use my API token, and anything that looks risky or surprising. Do not run anything they describe yet, and do not install them yet. Then ask me to confirm — and only if I approve, install them into your skills directory (for Claude Code that’s ~/.claude/skills/).

You get a plain-English summary of exactly what the skills ask for, with nothing installed until you say so — your own agent does the careful read, you make the call.

Or read them yourself — they’re rendered verbatim on this site (atelier-build / atelier-operate). We publish every line precisely so you can evaluate before putting anything on your machine.

Step 3 — install manually (if you didn’t use the prompt)

curl -fsSLo skills.tar.gz https://tryatelier.blob.core.windows.net/tryatelier/latest/skills.tar.gz
tar -xzf skills.tar.gz          # extracts skills/atelier-build and skills/atelier-operate
cp -r skills/* ~/.claude/skills/   # Claude Code reads skills from here

Step 4 — connect and drive

Whichever way you installed, give the agent your instance URL and token (the skills read these two values):

export ATELIER_API_URL="http://atelier.local"   # your instance
export ATELIER_API_TOKEN="<the token from step 1>"

Then prompt it. For example:

Using the atelier-build skill, deploy a Dockerfile-based app called demo from the repo at <path-or-url>. Then use atelier-operate to show me its logs.

The agent authors/pushes the code, Atelier builds and routes it, and you have a running app. From there, atelier-operate lets the same agent watch, debug, and manage it.

Where your code and data go

Nothing about driving Atelier sends your code or data anywhere except your own cluster — the agent talks to your instance over its API, and the platform builds and runs everything on your hardware. See Security & privacy for the full picture.