Atelier v0.25.0-beta

This release is about bringing more of your real code in — and seeing what happens after it deploys. You can now clone private repositories, import a repo to work on without building it yet, and when a deploy goes wrong, a new Health view tells you why without dropping to kubectl. Apple Silicon / OrbStack installs also get smoother, and there’s a round of reliability polish.

Coming from v0.24.x? Straight upgrade — one command, nothing you run changes, and everything here is additive. Your apps, folders, pins, alert channels, public-access config, and webhooks carry over untouched.

Clone private repositories

The Clone Repo flow could only reach public repos. Now you can clone a private GitHub, GitLab, Bitbucket, or self-hosted HTTPS repository: expand the new Private repository section in the Clone dialog and paste a personal access token (Atelier fills in the right username for your host automatically).

The token is used once, for that single clone, and then discarded — it’s never written to the database, the app’s repository, the build logs, or anywhere else. Once your source is in, Atelier doesn’t need the upstream credential again.

Import a repo without building it

Sometimes you want to bring a codebase in to work on rather than deploy it straight away — to add a Dockerfile, review it, or adjust it first. Tick the new Clone only (don’t build yet) option and Atelier imports the source into a new app and stops, leaving it ready to build. No Dockerfile is required at this stage; add or adjust one and push, and it builds and deploys like any other app. It’s also the clean way to bring in a repo that doesn’t have a Dockerfile yet, instead of hitting an error.

See why a deploy is unhealthy

When an app failed to come up — a crash loop, a bad volume, an image that won’t pull — the portal used to go quiet, and finding the cause meant kubectl. Each app now has a Health tab that shows it for you, live:

Rollout progress — how many replicas are ready vs. expected, so a stuck deploy is visible instead of silent.
Why a pod is unhealthy — the actual reason (CrashLoopBackOff, ImagePullBackOff, OOMKilled, a failed mount or scheduling), the container’s exit code, and restart counts, with problems highlighted.
The crashed container’s logs — one click to see the logs of the container that died, which is usually where the cause is.

This works whether or not you have a language model configured — it’s plain Kubernetes truth, surfaced in the app view. The crash-alert banner now has a View health shortcut straight to it.

Smoother on Apple Silicon and OrbStack

Installing into a Mac VM (OrbStack, Lima, Colima) could stall on storage, because the default storage layer can’t attach volumes in those environments. You can now install with ATELIER_STORAGE_CLASS=local-path to use the built-in local-path storage, which needs none of that — perfect for a single-node dev machine. (Keep the default for a real multi-node cluster.)

Reliability and polish

Public access starts more reliably. The Cloudflare Tunnel could fail to come up on a cold cluster if DNS wasn’t ready yet; that startup race is fixed, with a more recent tunnel image.
Leaner build-log fetches for agents. The build-log API can now return just the tail or just the errors, so an agent driving Atelier spends far less of its budget reading build output.
Updated agent skills. The bundled Agent Skills now cover private-repo clones, importing without building, and how to run a database alongside an app.

Upgrade

One command:

curl -Lo upgrade.sh https://tryatelier.blob.core.windows.net/tryatelier/latest/upgrade.sh
chmod +x upgrade.sh
./upgrade.sh v0.25.0-beta

It rolls atelier-core, atelier-ui, and the MCP server pods. Your running app pods aren’t touched, and all your settings carry over unchanged.

Fresh install on a Mac / arm64 VM: use the atelier-install-arm64 binary, and install with ATELIER_STORAGE_CLASS=local-path for an OrbStack / Lima / Colima machine.

Bug reports and feedback in the testers Discord channel. Thanks for testing.