Atelier release history
This page consolidates per-version release notes for older Atelier versions. The three most recent releases (v0.20.0, v0.19.0-beta, v0.18.0-beta) still live on their own pages. Anything further back is summarised here as a one-paragraph headline of what shipped.
Versions are listed newest-first.
v0.17.0-beta
Connecting agents and tools to Atelier. Durable, named, role-scoped API tokens mintable from the UI (Settings → System → API Tokens) — Viewer / Developer / Admin, with optional expiry and one-click revoke. Two portable Agent Skills (atelier-operate, atelier-build) shipped — Markdown bundles that let an external agent operate or build apps over the API, so a self-hosted assistant can manage the platform from a chat app.
v0.16.0-beta
Expanded what you could bring to Atelier as a ready-made container image — custom start command, persistent storage, multiple ports — so real off-the-shelf applications could run without writing a Dockerfile. Also hardened the platform around the in-cluster image registry so a full disk couldn’t quietly break builds.
v0.15.1-beta
Patch on top of v0.15.0. Lint now runs on Orchestrator-engine builds — v0.15.0 had the lint stage running only for the Classic engine; Orchestrator builds silently skipped it. Under Hardened, an Orchestrator-build finding raises an advisory lint hold after deploy (same shape as the post-deploy CVE hold). The four linter images (ruff / eslint / clippy / golangci-lint) now ship bundled with the release for offline/air-gapped installs.
v0.15.0-beta
A feature release. The Apps page becomes an organisable dashboard you can shape to your fleet — folders, pinning, manual order, search and status filters, layout saved on the platform and follows you across browsers. The build pipeline gained a code-quality lint gate (#388) — ruff / eslint / clippy / golangci-lint — with profile-dependent behaviour: Standard runs lint informationally and ships anyway; Hardened holds the deploy until you fix the code and redeploy, or click Override to ship that build anyway. Auto-deploy webhooks went end-to-end — Gitea’s defaults used to block deliveries to the in-cluster address (silent never-fires); fresh installs now allow it out of the box, and you can pick on a per-app basis what a push rebuilds with. Plus a Settings tidy-up and a little life on the login screen.
v0.14.1-beta
A patch release quieting three sources of recurring supervisor noise that v0.14.0 testers hit immediately. OOM alerts stop refiring once the pod has recovered — the detector was reading “this container ever OOM-killed” rather than “currently in trouble”; now a pod stable for 15+ minutes after its last OOM-kill stops being flagged. Universal-runner MCP servers (npm/pip packages) get more memory headroom — the catalog 512Mi limit was borderline at npm/pip install time; raised to 768Mi for npm-package / pip-package deployments (catalog servers like Fetch and Brave Search stay at 512Mi). Orphan-Deployment warnings now propose a one-shot cleanup action instead of re-alerting every 30 minutes for the same orphan. Plus a small docs addition that cuts the time to wire up a new MCP server.
v0.14.0-beta
The observability release. You could now see what the platform is doing in real time, get pushed an alert when something needed attention, and wire those alerts to wherever you want — Telegram, Slack, email, anything that speaks MCP. Plus a handful of platform fixes that closed gaps a fresh install would otherwise have walked into.
v0.13.1-beta
A patch release fixing a fresh-install regression in v0.13.0. The new installer step that auto-installed Longhorn’s host dependencies (open-iscsi, nfs-common) on Debian/Ubuntu ran after k3s was up — and apt’s needrestart hook could bounce the k3s API server at the wrong moment. Resequenced.
v0.13.0-beta
The onboarding release. A fresh install became simpler, more self-healing, and actually guided you to a first successful build instead of leaving you to discover the rough edges. If you had previously bounced off a stalled install or a build that failed with no LLM configured, this was the release that fixed that.
v0.12.0-beta
The observability release (first of two — superseded in narrative scope by v0.14.0). Atelier started watching its own health — the platform’s pods, its storage, its background processes — and surfacing it when something was wrong instead of failing silently. The orchestrator build engine also got materially more capable, and a couple of long-standing rough edges were closed. Upgrade path from v0.10.1 or earlier started being handled automatically across the v0.11.0 MCP transport migration.
v0.11.1-beta
A tiny patch on top of v0.11.0. Two fixes that surfaced during tester rollout: fresh installs could now deploy brave-search cleanly (the catalog used to reference an in-cluster registry path that didn’t exist on fresh clusters), and a related deploy-flow gap was closed.
v0.11.0-beta
The MCP transport migration. Atelier started speaking the modern Streamable HTTP transport defined in the MCP spec (2025-03-26 onwards), replacing the deprecated HTTP+SSE pattern. Closed the silent crash-loop the mcp-fetch and mcp-brave-search pods had been carrying for ~2 months (105 and 131 restarts each in tester clusters) under concurrent use, and laid the foundation for future agent-app MCP integration.
v0.10.1-beta
Orchestrator-mode parity polish. The update path now ran through the orchestrator when appropriate (previously hardcoded to classic even for orchestrator-built apps), and orchestrator-built apps were picked up by the base-image monitor.
v0.10.0-beta
The big one. A new orchestrator-mode build engine shipped alongside the existing pipeline as opt-in — a tool-calling LLM agent that figures out the build path itself, rather than going through hand-coded stages. Plus a skill registry that captures stack-specific platform wisdom across builds, plus live cost visibility on every build regardless of which engine you pick. Three weeks of spike work (#392) condensed into a release.
v0.9.10-beta
A bug-fix release. Polish on the staged review-and-revise flow shipped in v0.9.9-beta — a few rough edges that surfaced once people started actually using it.
v0.9.9-beta
A small, focused release. Review-and-revise for Upload Local, Clone Repo, and From Gitea flows — previously these jumped straight to a build with no review step; now they match the chat workspace’s planning loop, letting you refine what the LLM intends to build before clicking Build.
v0.9.8-beta
A platform-distribution release. Fresh installs became a single binary download (platform images now pull directly from ghcr.io/atelier-project/... on first boot, rather than needing three image tarballs alongside the installer). In-app version awareness improved, and the public docs site went live.
v0.9.7-beta
A polish release focused on observability, control, and pipeline correctness. The headline change: the “Ask Nova →” button on the error banner finally did what the original intent had been — packages the failure context (platform error + last 50 build-log entries) and routes it to Nova for analysis.
For a complete list including the three most recent releases (still standalone), see the release notes index.