Atelier v0.21.0-beta
Atelier sharpens its focus and grows an agent of its own. The platform now builds, deploys, and operates the code you bring it — yours, or your AI agent’s — and no longer writes application code itself. To fill that gap it ships an optional build agent that lives in your cluster: open a terminal, describe the app you want, and watch it write the code and ship it through the same pipeline as everything else. And the whole interface got a fresh coat of paint.
Coming from v0.20.x? Straight upgrade — one command. This is a beta: the new in-cluster build agent (its deployment and persistent storage) is getting a soak before the next stable tag. Nothing you already run changes — the agent is off until you turn it on.
Build with AI — in your own cluster
Turn on the bundled agent in Settings → AI → Bundled Agent and a new Build with AI terminal appears. Describe what you want in plain language; the agent scaffolds the app, writes the code, pushes it, watches the build, and hands you the URL when it’s live — the same path you’d drive by hand, done for you.
Pick your agent. Choose Claude Code or opencode — the same terminal, your choice of driver:
- Claude Code — paste an Anthropic API key when you enable it. The key is stored only inside your cluster, never in the platform database.
- opencode — no key up front. The first time you open the terminal you run
/connectto sign it in to your provider (Anthropic, OpenAI, a local model, and more). The sign-in is saved on the agent’s storage and survives restarts, so you only do it once.
It’s yours, attributed to you. Each terminal session gets its own access token scoped to your role, so apps the agent creates show up as created by you. Close the tab and the agent keeps working; reopen and you’re back in the same session. Your provider, your bill — no markup, no per-seat fee.
Bring a different agent. Prefer your own? Any skill-aware coding agent — Claude Code, Cursor, a chat assistant — can drive Atelier from your machine through the same tokens and skills. And you can swap a different agent image into the in-cluster terminal entirely.
A more focused platform
Atelier is now squarely a deployment platform. It takes working code — from you or from an agent — and turns it into a running, operated app on your own cluster: container builds, generated Kubernetes manifests and routing, CVE scanning, deploy gating, logs, metrics, and crash diagnosis. What it no longer does is author your application code. The AI features that remain — the ops assistant, crash diagnosis, build summaries, and code review — are analysis only, and every one of them is optional.
A fresh look
A visual refresh across the app: a new palette and self-hosted typography, a reworked Apps board with cleaner cards and status pills, a refined navigation rail, and a tidier per-app view that trades its old side strip for clear horizontal tabs and a calmer header.
Smaller things
- Clearer LLM settings. The LLM Profiles screen now explains what it’s for up front — the platform runs fine with no LLM at all; profiles only power the optional AI features — and spells out which model runs each task.
- Docs and website refreshed to match the new focus.
Everything still works with zero AI keys: push your own code, clone a repo, import an image. Add a provider — or turn on the build agent — when you want the AI on-ramp.