v0.6.6

Fixes a regression present in 0.7.1 where successful workspace scripts reporting Exited with code 0 could fail deployment (#832). Single-image and Compose builds now distinguish Docker Engine errors from application output, for both classic and BuildKit builders. Diagnostic hints enrich real failures without overriding successful completion or hiding Docker's error; SSH builds continue to use the command's exit status. Cancellation and inactivity safeguards remain in place.

Settings → Git now supports one-click GitHub.com manifest creation and manual GitHub/GitHub Enterprise Server registration, without an Openship Cloud or OAuth dependency. Private keys, client secrets, and webhook secrets are encrypted at rest; source-specific JWTs mint short-lived repository-scoped tokens against the correct API origin; GitHub sends verified lifecycle, push, and check webhooks directly to the instance. Owners can install an App on multiple accounts, rotate and verify credentials, select a default source, and safely delete one while project bindings move to an available replacement. Existing environment-backed Apps, gh, personal tokens, and per-server SSH keys remain compatible fallbacks (#782).

Install callbacks bind a durable one-shot nonce to the initiating user and workspace and verify the reported installation against the configured App before atomically persisting and reconciling it. The cloud callback survives API replicas/restarts, rejects removed workspace members, and cannot bind another tenant's caller-supplied installation id. Installation rows, token caches, project bindings, lifecycle webhooks, and reconnects are workspace-scoped; webhook arrival order never guesses a workspace from membership order (#759).

The claim page loads safely before sign-in, routes existing users back after login, offers token-bound account creation only when authorized, and exposes a copyable link beside each pending invitation. Invitation authorization, identity, credential, personal workspace, and membership now commit under one locked transaction, so cancellation races and failed signup cannot leave an unauthorized or partial account that blocks retry. Accept, reject, cancel, and signup mutations are serialized per invitation across API replicas. API access logs and Edge request telemetry also redact invitation bearer tokens and discard query strings before storage (#743).

Saving a service route creates the same pending domain row as every other custom domain, then the shared records panel previews provider changes and applies them only when requested. The API documentation no longer promises a silent provider write that the product deliberately does not perform (#760).

Server readiness no longer inherits the control plane's DEPLOY_MODE. Docker and Git are checked for every managed target, while one shared dependency planner inserts and orders Docker ahead of the container-based Edge across streamed setup, SystemManager, and direct component installation guards (#767).

Top-level project variables are recovered from the trusted source, encrypted into each deployment snapshot, and passed through the existing shared Docker build-argument resolver for first deploys, redeploys, webhooks, folder uploads, and explicit single-app mode. Native services[].buildArgs adds per-image overrides (null inherits the matching project value), while service env remains runtime-only (#795).

Docker's reverse h2c session now bypasses Bun's incompatible node:http upgrade path while retaining dockerode's BuildKit session and registry-auth service. Local Compose builds using syntax directives, cache mounts, heredocs, and other BuildKit-only Dockerfile features no longer fail on /session before the Dockerfile starts (#745).

Dashboard and desktop update surfaces now read the matching version from the tagged CHANGELOG.md instead of repeating the GitHub Release body. The native update panel keeps advisory copy separate from a resizable, scrollable changelog, and all changelog actions open the version page on openship.io rather than GitHub (#752).

Resolving the tracked branch head refreshes its upstream cache while a deployment is in progress, and successful deployments invalidate cached upstream state before the UI refreshes. The update banner and issues feed therefore compare the live deployment against a fresh upstream version instead of an older cached commit or image digest (#726).

An explicit port-only choice now persists a fixed all-interface Docker binding, including for catalog routes such as Supabase Kong that declare a served port but no Compose ports entry. This makes the wizard's server URL and {{publicUrl:…}} substitutions truthful. Retrying a failed draft reconciles the binding idempotently, and switching to a domain removes or restores only the mapping owned by the installer (#770).

The final setup step combines the target's listener table with bounded off-box TCP checks for 25, 465, 587 and 993. The Health tab shows each port as a first-class signal, reusing a coalesced 60-second sweep, and test-email timeouts now identify provider firewalls/security groups/security lists/NSGs instead of blaming a stopped SMTP daemon when Postfix is demonstrably listening (#755).

Synthetic ULA AAAA responses from Clash/sing-box are now unknown, matching the existing A-record behavior. Together with the informational SpamAssassin row and normalized split-string DKIM comparison, this closes all three false-positive paths reported in #240.

Webmail continues to submit to the local Postfix on 465, which alone owns relay scope, DKIM signing, and provider credentials. The identical timeout reported in #391 now receives the shared public-port diagnosis; provider relay credentials are never copied into the webmail container.

Cached openship@<domain> credentials are reused only while the live mailbox and forwarding rows remain active. A deleted row is recreated automatically, and an already-cached transporter rotates once and retries safely on an auth rejection (#754).

Ordinary mailbox edit and delete operations reject the Openship-owned sender, while the Mailboxes tab exposes a dedicated password rotation/repair action that also works when the row is missing.

Relay maps and additional-domain DKIM configuration now elevate only root-owned host-file operations while Docker commands keep the login user's daemon context. Relay credentials are published atomically as 0600, and a server without root or passwordless sudo receives an actionable typed refusal unless the mail bind mount was explicitly made writable for that login (#756).

Cancellation propagates through post-build preflight, human prompts, host-scoped provisioning locks, activation, health checks, routing, and both single-service and Compose pipelines. A durable database signal also wakes a worker running in another API process.

A cancelled row keeps its execution lease until cleanup has finished, while the dashboard shows the cancellation still stopping instead of offering a redeploy that cannot safely start. The API and CLI report success only after that lease is released; otherwise cancellation remains explicitly pending. Refreshing reattaches to that server-backed state (#757).

The cancelled transition is conditional and atomic, and cleanup excludes containers carried forward from the active Compose release. If ownership cannot be proven, cleanup fails closed.

Removing one or more environment keys from Compose now preserves the running service configuration and creates an explicit drift review, even when the operator never edited the imported baseline. Accepting that review remains the deliberate way to apply the deletion.

project create --local-path ... --type services now uses the API's canonical scanner and persists the detected rows immediately. Deploy and redeploy also re-read the same local source through the existing three-way reconciler, so a legacy zero-row project self-repairs instead of entering an empty services pipeline (#751).

Raw manual syncs preserve ${VAR} provenance and resolve it from the final project/service environment at deploy time, while Docker Compose-normalized CLI syncs carry an explicit final-value marker so escaped $$ is never expanded twice (#751).

The deployment lifecycle writes one deduplicated terminal error into the durable build log for every lifecycle failure path, including an empty Compose service set, instead of leaving the explanation only in deployment.error_message (#751).

project create --server <id> stores the initial binding, while deploy --server <id> drives Git and folder uploads through the same org-scoped target resolver and deployment preflight as the dashboard. The deployment response exposes a stable id for --watch, project get shows the resolved target, and a failed deploy never retargets the project (#763).

Before changing containers or routes, a normal deployment checks the target for an untracked Docker Compose stack with the same project slug and service names. A collision now stops with FOREIGN_COMPOSE_STACK and tells the operator to import/adopt, remove, or rename it. Explicitly adopted containers and containers already tracked by Openship continue normally.

A routed service keeps its active host port on its current physical server. If that port cannot be reclaimed safely, deployment stops before routing changes instead of allocating a new port and leaving the public hostname pointed at the wrong workload. A real server migration may still allocate a new host port.

Normal projects no longer claim same-slug containers through Docker Compose labels alone. Native Openship labels, the canonical managed name, or the tracked container id establish ownership; Compose-label recovery remains available only for an explicit adoption.

Adding environment/build-argument template provenance to an older service baseline is treated as a one-time metadata upgrade, not as a Compose edit. Real changes arriving with that metadata are still reported, and operator-edited live values remain intact while the baseline advances.

The dashboard adds a Health view for every tracked project service, with fleet counts and state charts for healthy, unhealthy, crash-looping, down, and unknown workloads. Rows show the serving server and last observation, and link back to the service that owns the container.

The page reads an in-memory snapshot produced by the grouped health sweep and Docker-event accelerator; it does not create a poller per container or perform Docker I/O on page refresh. Per-service monitoring can be disabled before the workload enters the watch, inspection, event, or notification path.

Bulk component updates start together without holding a modal open, retain their durable sessions across navigation and refresh, and allow an operator to reattach to a specific log. The checker batch similarly exposes one shared progress session so multiple tabs cannot multiply Git, DNS, Docker, or version scans.

The destination creates a single-use, ten-minute receive code for Replace or Merge mode; the source encrypts the selected database rows and plaintext credential bundle directly to that destination, and the destination immediately re-encrypts every transferred secret under its own instance key. Server SSH passwords and keys, environment variables, tokens, registry and DNS credentials, and backup credentials move without sharing either instance's BETTER_AUTH_SECRET or managing a transfer passphrase (#656).

Settings now shows the durable row count and a count beside each optional history group. Analytics, audit/notification activity, backup history, incident history, and migration history can be included independently; configuration, projects, services, servers, users, and credential records always stay in the portable core. Leaving the selection absent preserves the legacy full export (#656).

The offline export/import path remains available, but a file containing a sealed credential bundle now requires its passphrase before any database write. Invalid bundles are rejected before restore rather than importing rows whose secrets were silently scrubbed.

A project with composePath is materialized into the service pipeline instead of falling through to one generic Dockerfile build. Map and list forms of build.args, bare arguments, and ${...} expressions are stored per service and survive CLI sync, reconciliation, migration adoption, redeploy snapshots, and rollback. Docker socket, SSH, batch, and cloud builds all use the same argument resolver, and Openship's build-command logging does not print argument values (#689).

An empty args map or a deleted args key clears stale values and interpolation provenance, while snapshots created before build arguments existed remain non-destructive during rollback (#689).

Malformed arguments, repository-escaping or remote contexts, and build features Openship cannot reproduce no longer leave the old service shape running silently. A declared Compose project with no materialized rows is scanned once to bootstrap its topology, while an explicit single-app choice remains single-app.

Raw expressions such as postgresql://user:${POSTGRES_PASSWORD:?set it}@postgres:5432/app are evaluated against the final project, frozen-release, inline, and service-scoped layers. Embedded requirements, passthrough keys, defaults, nested expressions, and escaped dollars retain Compose semantics; an unresolved required variable fails that service with the missing key named (#673).

The service Environment tab now owns service-scoped environment rows rather than rewriting the Compose-owned service.environment object. Values added in the UI therefore survive a Compose reparse and project redeploy. Saving an unrevealed secret — including renaming its key — preserves its existing ciphertext instead of storing the mask or deleting the value.

A single-app project can resolve versions from GitHub Releases or an HTTPS version feed, optionally pin a version, and render that tag into a registry image template. Deploy and Update pull the application image directly on Docker or Cloud — no Git clone, Dockerfile build, or release archive — while drift detection compares deployed and upstream semver. Successful Docker releases freeze the immutable registry digest so rollback can reacquire the exact image after local retention expires; source changes apply atomically across environments and are available in both the dashboard Source tab and openship project release-image (#694).

Docker and host-mode projects reuse the active deployment's retained artifact even though they have no service rows. Apply never fetches Git or silently turns into a rebuild; if the artifact is gone, it says to Redeploy instead (#674).

Pressing Deploy now shows records for single-app custom domains, service scalar routes, and multi-route Compose endpoints. Multiple hostnames appear together, www is grouped with its apex, and a remote deployment preview uses the selected server's public address rather than this instance's address (#663).

Deployment reconciliation preserves both pending and verified custom-domain rows and their live targets when a release omits them. A failed deployment rolls back only generated managed routes, never user-owned custom configuration, and concurrent hostname claims use database-authoritative ownership instead of letting the losing project route another project's domain (#675).

Loopback ports now have database-enforced, physical-target-wide ownership across organizations. Allocation reserves the port before Docker binds it, treats local and “This Server” as one namespace, and persists every routed Compose port through stopped and reconciling states. A stale vhost/TLS certificate can therefore never be repointed accidentally when a later deployment starts listening on the same number (#682, GHSA-284v-9jw3-jfhx).

When openship-edge is stopped or missing, Retry Routing reconciles and health-checks it before touching vhosts. An unrecoverable edge returns an actionable warning immediately instead of hanging until the route request times out (#693).

Route cleanup validates the installed configuration and signals only the verified running master instead of starting a second nginx process on ports 80/443. Listener checks account for every socket owner, including mixed SO_REUSEPORT listeners, and fail closed when ownership cannot be proven. Delete flags now work consistently in query strings and JSON bodies; forceOrphan records unfinished host cleanup durably before removing the project so the garbage collector can retry it safely (#700).

Preflight and the build pipeline now share one source-location plan. Local socket and TCP daemon builds prepare source on the API host, remote SSH Docker builds may clone on the target, and bare or cloud builds retain their own boundaries. A local server row therefore no longer asks a nonexistent remote clone path to use ambient Git credentials (#654).

Bind sources are canonicalized on the machine that owns the Docker daemon, so /var and /etc resolve to their /private/... targets before Docker Desktop or OrbStack sees them. A healthy-looking edge with stale logical mounts is recreated onto the same vhost, certificate, ACME, and static-data directories (#692).

Fallback token discovery reads exactly the hosts.yml selected by GitHub CLI precedence on Linux, macOS, and Windows. If GH_CONFIG_DIR or XDG_CONFIG_HOME is set but missing or tokenless, Openship no longer falls through to another user's home directory and borrows that credential (#687).

Every mail-engine boot now refreshes resolv.conf and the supporting NSS files inside the persistent Postfix spool before the supervisor starts. The engine fails closed if no resolver can be installed, rather than starting an SMTP service that rejects every legitimate sender with 450 Helo Host not found (#686).

Creating an environment commits it to shared state immediately and reconciles the canonical list; deleting one removes it from every affected cache and navigates to the best surviving sibling. A failed follow-up read no longer makes a successful create look like it failed (#657).

This release resolves a number of reported security issues, with regression tests added for each. Areas touched: authorization on instance-wide and GitHub operations, access-token minting, path handling, the mail engine, webmail's HTML sanitizer, the desktop shell, and redaction in stored build output. Details are published as advisories on the repository. Upgrading is recommended.

Host-control operations are pinned to the host channel rather than defaulting to the API process's own container, cloud builds can never run on the API host, and stored SSH key material is encrypted at rest and stripped from any cross-host export.

When Openship runs in Docker, a handful of operations are genuinely host-level (freeing a foreign proxy off :80/:443, host system config, the mail engine, writing a catalog app's generated config), and they reach the host over SSH on the internal bridge. Install now generates the key, appends a restricted from=-pinned authorized_keys line, checks that sshd listens on an address containers can reach, opens the port in whichever firewall the host runs, and verifies the round trip — instead of reporting success and failing on the first host operation weeks later.

This address is host-local, so it traverses filter/INPUT where a default-deny ufw lives; published container ports are DNAT'd and skip it, which is exactly why the rest of the stack looks healthy while this one channel hangs. Firewall, sshd listen address, key, and reachability are each probed and reported separately.

The CLI preflight, the API's boot banner, openship doctor, the dashboard's server banner, and the deploy log a host operation dies in had drifted into different stories. They now read from one shared vocabulary, because every line is either "your install is fine, this one feature isn't" or "your install is broken" — and an operator who reads the wrong one either tears down a working box or ignores a dead feature for months.

An install from before this work, or a raw docker compose install, can be fixed in place by updating, rather than re-running up and risking the environment. A deploy that needs the channel emits the notice once per deploy, before the fan-out, instead of once per service.

An operation that needs a root-owned path fails with the actual cause instead of mkdir: cannot create directory '/root': Permission denied.

.env.example spells out all five steps (key, restricted authorized_keys line, sshd listen address, firewall rule, recreate the api because env_file: is read at container creation), and a new troubleshooting page walks the repair.

Docker installation was curl get.docker.com | sh on every Linux, a script that hard-refuses Amazon Linux, AlmaLinux, Oracle Linux and Alpine. Each call site had its own package-manager table (there were five, and they disagreed), its own systemd test, and its own reading of "host outside my allowlist" — always "nothing to do". Host facts now come from one detector, the commands that act on them live in one module, and adding a distro is a compile error rather than a silent no-op.

Only an explicit reinstall overrides an engine that's already running, and an installer that skips an already-working component says so instead of leaving the one actionable line nowhere at all.

Three layers answered "may I do root-owned work here?" differently: component installs gated properly, toolchain installs never gated at all, and the server state store assumed the login was root. There's now one privilege resolver, and it elevates the write rather than the verify (so a sudo shell never resolves a different $HOME). A forced-command host that can't report a uid is no longer told to "connect as root", which was the opposite of its fix.

The toolchain catalog and installer were rewritten onto the same host profile, so a bare-metal build on a non-Debian box gets its runtime installed instead of a not found at build time.

A new top-level command: edge up stands the proxy up and serves :80/:443, edge migrate takes over an existing nginx/Apache/Caddy and imports its sites, edge takeover and edge free claim the ports, edge sites lists what could be imported, and edge repair diagnoses why it isn't serving (--fix resolves a port conflict).

edge domains add app.example.com --port 8080 registers a hostname against a port and issues a certificate; edge rules manages per-route rate limits, bans and geo/CIDR access; and edge traffic, edge analytics and edge logs read what the proxy already records. Host operations are Linux-only; the control-plane half works from any machine, including desktop, against whichever context is active.

A new The edge guide for the tasks, and a full command reference.

A self-hosted project silently inherited the cloud free tier (0.5 vCPU · 512 MB) and OOM-killed memory-hungry images. On your own box the container's real ceiling is the machine, so 0 (no limit) is now the default and every consumer tests for a limit before applying a cap.

Caps were validated against hardcoded constants (4 cores / 8192 MB), so a 64 GB box could not be told to give a container more than 8 GB. The ceiling is read from the Docker daemon's own /info (NCPU, MemTotal) — identical for the local socket and a remote daemon over the pooled SSH bridge — falling back to the OS when Docker isn't reachable. Cloud is unchanged: a metered workspace is still sized from the tier table.

A cap larger than the host can allocate is rejected before any build work starts, with the machine's real numbers in the message.

An instance can present itself as Openship Mail (OPENSHIP_PRODUCT, or a toggle in Settings): the left rail becomes the mail control plane — the ten admin tabs promoted to nav entries across three headings, plus the host and settings rows an operator still needs — and the platform nav is hidden. This is presentation only, never an authorization boundary: webmail still deploys through the ordinary project pipeline, so platform endpoints stay live. Cloud is always the full platform.

The mail surfaces are scoped to a selected server rather than assuming one, with a switcher in the shell and the scope carried through the admin tabs.

The relay code carried a "ses" | "custom" union, so SendGrid, Mailgun and Postmark lost their identity the moment they were saved: no SPF include, no round-trip in the UI, and adding a provider meant editing an if in the service, the DNS builder and the scanner. Per-provider facts are now data that all three read, and split delivery (receive here, send through someone else) is a first-class setup.

The daemon sweep reports that nine processes are running, which says nothing about whether Postfix can hand a message to the next hop. One wrong character in a relay password gave nine green daemons, green DNS, a Test-tab email that reported success (our Postfix accepts and queues before the relay hop), and mail that quietly deferred forever. Outbound delivery is now probed and surfaced on its own, with a reworked Health tab and a summary that names the failing hop.

Live engine output in the dashboard, so a setup that stalls can be read rather than guessed at.

The mail admin Backup tab gains a real schedule and retention plan, wired to the same pruning the rest of Openship uses.

A preflight checks the things that used to surface mid-stream (password shape, firewall step, DNS), and setup errors render in a banner instead of vanishing into the stream.

It installs through the same generic installer, image, volume and routing pipeline as every other app, and installs from /apps just as well as from /emails. The mail-specific part is the only part that stayed: which IMAP/SMTP backend it belongs to — one of your Openship mail servers, or an external provider.

The old lookup read webmail-<serverId> back out of the project slug, which the generic installer never produces, and which mislinked any project someone happened to name "Webmail Prod". It's a real foreign key now, nulled (not orphaned) when the webmail project is deleted, with a legacy-slug fallback that stamps the column the first time it resolves a pre-existing install.

A dedicated webmail Dockerfile and a bootstrap script, so first boot provisions its schema instead of depending on the mail engine's.

Meilisearch, Redis, PostHog, Umami and Neon (Postgres), each declaring its internal connection so it can be linked into a project over the internal network.

Verified, unverified and hosting-mode badges with explanatory tooltips, so a community template isn't visually indistinguishable from one Openship has booted and checked.

The published app.schema.json gains the fields behind the above, and the reference docs and "Add an app" guide were rewritten around them.

registerRoute replaces the whole vhost, so a caller that omitted cleanUrls / redirects / headers didn't leave them alone, it deleted them. Those fields were built inline in two places, so a plain deploy neither applied them nor preserved them, and an unrelated redeploy silently erased whatever a "Retry routing" had installed. They're now compiled once, in a pure module every route-registration site carries.

A single resolver owns the proxy_pass target: a pinned loopback host port by default (stable across restart, never internet-facing), the container's bridge IP as an advanced option, and a transparent fallback to the container IP for an internal Compose service that publishes no host port. Selectable per instance, with auto as the default.

The redirect installed for a hostname is no longer reverted the moment certbot succeeds, and it no longer re-fires on internal rewrites.

The real-IP configuration is derived in one place for every scope, so free subdomains and custom domains agree on what the visitor's address is.

Per-route rate limits, bans and access rules are covered by an e2e test that drives the real proxy.

Connect a Cloudflare token and Openship writes the records a domain needs itself (#37), issues wildcard certificates over the DNS-01 challenge, and can route a domain through a tunnel instead of an open port.

The provider claims a record it wrote and leaves everything else in the zone alone, and it refuses to write a hostname it hasn't claimed. Deleting a domain in Openship therefore cannot remove a record that was already there.

The page and the modal were two 400-line forks that had drifted; they're now one component with a variant, so a field added in one place exists in both.

ssh_key_path points at a file on the API host, which is useless on a remote or VPS instance where your key lives on your own laptop. A key can now be pasted or uploaded from any browser, stored encrypted at rest, never serialized back to the client, and preferred over the on-disk path when present.

Reachability, host channel state, and what each one does or doesn't affect, rather than a single red dot.

The instance default target is a server or cloud; "this machine" is derived (on a server-host box it is already in the server list), so it's no longer an option that means two different things.

A rename modal, plus copy-id and pause/resume, on the heading where the name actually is. The slug stays immutable (it's infrastructure identity); the display name is free.

Pause, restart and logs run through the same deployment-runtime path as everything else, so they can't disagree with the platform about which host or runtime a service lives on.

Record-only deletes, orphan GC scheduling, pending actions and port checks were tightened around the same runtime path.

Two boolean flags (hasBuild, hasServer) each stood in for two different concerns, so a deploy could be accepted and then fail a stage later. A project is now described by three orthogonal axes — where the code comes from (git, a prebuilt image, or an upload), how it's built (Dockerfile, buildpack, static, or nothing to build), and what runs afterward (web, worker, or static) — resolved in one place that every stage reads. Existing projects classify exactly as they did before.

A Dockerfile build runs no buildpack step, so it naturally carried hasBuild=false, and the one gate that decides whether to fetch a clone token keyed off that flag — dropping the token and failing the clone on a private repo, even though the Dockerfile needs the repo as its build context. Whether the code is fetched now depends only on where it comes from, never on how it's built.

A long-running process with no port (a queue consumer, a bot, a cron loop) had nowhere to go: called a server it failed preflight for a missing port, called static it was routed through file-serving and failed for a missing docroot. worker is now its own workload — built like any other app, run as an always-restart container with no port and no route, and checked for a start command instead of a port. Choose it from a project's Server / Worker / Static switch, or set workload: worker in openship.json.

A variable you set by hand is kept and marked as yours, instead of being regenerated out of a fixed key list.

up detects that it would mint a secret an existing install already has (which would leave encrypted env undecryptable), and refuses rather than silently replacing it. The replaced .env is kept for recovery.

An explicit --image-version wins over the environment and the CLI's own version, for the case where a release reaches npm ahead of its images reaching the registry.

The volume's contents are probed before a cluster is started on top of them.

The host channel, an unprovisioned install, and port conflicts, in the same run.

A migrated 5432:5432 would bind host 5432 on the next deploy and collide with whatever already held it (often Openship's own Postgres), aborting the deploy. Adoption keeps only the container port; the service is reached by name on the project network and re-exposed from the Domains tab. The warning tells apart a port that was published off-box — Docker's publish rule routes past the host firewall, so it was genuinely reachable from the internet — from a loopback-only one, so you know whether any external reach was actually given up.

A service that asks for a longer shutdown window now gets a graceful stop, with its own signal and grace period, before it's replaced or torn down, instead of being SIGKILLed mid-write by an immediate force-remove. A container that doesn't opt in still stops the fast way, so there's no added redeploy latency.

The only way to let something run a command was a custom-command job, whose grant reaches every server in the org and can't be narrowed. Command execution now hangs off each resource's own permission: a grant on one server (server:admin) is a host shell confined to that box, and a grant on a project (project:service:write) runs commands inside that project's service containers — nothing wider. Every run is audited with its command, working directory and result, and container exec is refused on a runtime that can't isolate it.

Two new MCP tools expose host and in-container execution, each requiring the matching per-resource grant, with a hard timeout and an output cap (the agent's call blocks on the reply and the whole body lands in its context). Tools that only work on a self-hosted box are now hidden on the hosted platform, instead of being advertised there and returning 404.

The MCP consent screen, personal-access-token scoping and member grants were three separate editors and are now one. A connected agent's access can be edited in place from Settings — narrowed, or widened (widening to unscoped asks first) — and applies on its next request, with no disconnect-and-reconsent.

Several platform areas (jobs, notifications, analytics, settings, updates, cloud) could be checked for but never handed out, so there was no middle ground between "deploy only" and "the whole account"; they're grantable now, with billing and audit marked sensitive. Listing a collection honors a wildcard grant too — a token allowed to read a server by id no longer 404s when it tries to enumerate them.

Desktop mode is built to control remote servers; running the workload on the desktop itself isn't enabled yet, and the UI now says that once, in one place, instead of offering a target that fails later.

See Security above.

"The edge" guide, the openship edge command reference, host-channel troubleshooting, a rewritten "Add an app" guide and app catalog reference, plus notes on installation, updating and logs/monitoring.

Generated OG images per page.

Entries collapse to headlines and expand to detail, with controls to expand or filter, driven by this file.

An instance with no sign-in form grants its session only to a browser on the same machine, so bouncing a remote one was guaranteed to fail, leaving a spinner and then either a connection error for a host that isn't yours or an endless redirect. The page now names the cause.

Pruning no longer considers a mail engine backup an ordinary project artifact.

A same-server migration attached a container that was never published, leaving it unreachable behind a verified domain.

Cancellation reached some paths and not others, so a cancelled build could still be deployed: remote Docker builds kept running, and the static extract went on to publish the output of a build you had already stopped. A cancelled deployment is now terminal.

A redeploy where every service was carried forward unchanged settled without advancing, so it no longer displaces the release that is actually running.

Editing project settings preserved the healthcheck test rather than dropping it, and a NONE test array reads as disabled instead of as a command.

Exec-stream sinks close on EOF, and the upload stall bound matches the producer's idle budget, so a completed backup can't wait forever on a stream that already ended.

Projects that rely on the image's own entrypoint deploy without one.

The release's new copy landed across all shipped languages, with the parity test extended to the new namespaces.

Every project (a plain single-app, a raw Compose stack, a monorepo, or an imported one) now shows a Connection card with its internal address (http://<alias>:<port>) and can be picked as a "Use in a project" source. Previously only catalog apps with a declared connection block, or multi-service Compose stacks, were internally reachable; a single-app project had no internal address at all. Same-boundary reachability is automatic and crossing a boundary stays explicit — an alias only becomes reachable once you link a consumer onto the network. Self-hosted only, and skipped for static sites (no listening port) and cloud-hosted apps (which link over Public instead).

A service can carry a custom east-west DNS alias that resolves alongside its default name, set from the service's Settings tab and available to single-app, Compose, and monorepo services alike. Openship normalizes it to a valid hostname and rejects one that collides with another service on the same project network, and the internal-address card shows the alias its containers actually answer to. Self-hosted only.

The service Settings tab groups fields into labelled sections and swaps free-text boxes for structured editors: ports and volumes are entered as chips (with per-item limits enforced as you type) and depends-on is a picker of the project's other services. The image and build-from-source fields are shown together — fill either, and a build context wins over a stale image — instead of an image/build mode toggle.

Until now a container was health-checked exactly once, during the ~15s window right after a deploy, and then never looked at again unless a human was staring at the project page. A new health watch polls every server's Docker daemon once a minute (and reacts within about 10 seconds when Docker itself reports a container dying or restarting), so an app that falls over or drops into a crash loop at 3am raises an alert through your existing notification channels instead of waiting for a user to complain. Self-hosted only.

A container stuck in a restart loop is broken on every poll, so a naive watcher would fire 60 alerts an hour about one fault. Instead each fault becomes a single incident: one message when it opens, one more only if it gets strictly worse (unhealthy → crash looping → down), and one all-clear that tells you how long it was down. A fault must be observed twice before it notifies, and an in-flight deploy, an operator's own docker stop, or a disabled project never trips a false alarm.

When a box's Docker daemon stops answering, Openship opens a single server-level incident that names how many projects it can no longer monitor, and freezes those projects' health rather than falsely marking them recovered. The record is durable, so a box that has been down for days does not re-page every time the control plane restarts.

Each project gets a Health tab listing its currently open incidents plus a rolling 30-day history of resolved ones with downtime durations. It also states whether monitoring is actually switched on (an operator toggle in the Jobs module) and surfaces the server's own outage when a box is unreachable, so an empty tab is never misread as "all good." Self-hosted only.

The Monitoring tab draws a per-country choropleth of visitors alongside a ranked list, and animates real-time request ripples fed by the same stream the Logs tab tails, so you can watch where requests are landing right now. It works the same way self-hosted and on Cloud, and it can tell "no visitors yet" apart from "geo lookups aren't set up on this box," which used to render identically as an empty map. On a multi-domain project the whole view scopes to a chosen domain.

The dashboard's visitor figure was a request count relabelled "unique IPs," so five page views from one browser counted as five people. It's now a genuine count of distinct visitors, deduplicated at the edge with a per-day salted hash that never stores an address or a per-visitor row — the number is honest while still collecting no behavioural data on your end users.

The Top Paths table was hardcoded empty and error rate couldn't be answered from anything saved. Both are real now: the daily rollup records normalized request paths (query strings stripped, ids collapsed, so tokens never become keys) and the full status-code mix, and a new card expands each 2xx/3xx/4xx/5xx class to the exact codes beneath it. On self-hosted, per-path aggregation is opt-in per project (a toggle, default off, including on existing projects) because it adds measurable work to every request the edge handles; on Cloud paths are aggregated automatically. Every other analytics dimension is unchanged.

CPU, memory and network usage used to exist only as a live stream that vanished when you closed the tab, so "was memory climbing before it got OOM-killed at 3am" had no answer. Usage is now sampled on a schedule into 5-minute buckets and charted over time, per service. A Compose project reports its whole stack with a per-service breakdown and an All/service scope picker instead of just the one container the primary service owned; a static site (nothing to measure) now says so plainly instead of showing four zeroed metrics. Self-hosted Docker deployments; bare non-Docker deployments report zero network, since per-process accounting isn't available there.

A domain pushing more than about 2 GB in a single minute (roughly 286 Mbps — routine for video or large downloads) used to overflow a 32-bit byte counter and kill the analytics scrape for that entire domain, so the busiest sites were precisely the ones with no analytics. The bandwidth counters are now 64-bit and collection keeps up. Self-hosted edge.

Traffic analytics live in the edge's memory under a TTL and were only flushed to the database when someone opened a project's Analytics tab, so a project nobody viewed for a day lost its per-minute data. A background scrape job now persists every server's analytics on a schedule whether or not anyone is watching, collecting a whole server in one pass instead of a separate memory scan per domain.

Tooltips were effectively transparent in the dark themes (plot lines showed straight through the text) and on light cards the hover marker rendered as a stray white block below the tooltip. Both are fixed with a shared tooltip style reused across the monitoring and billing charts.

A project's Routing settings now expose a full set of proxy tunables instead of a handful: raise or entirely remove the upload size limit, set connect/read/send/keepalive timeouts, control response buffering and gzip level, turn on HTTP/2, tighten request-header limits, and verify TLS to the upstream — each value validated before it reaches the generated nginx config and labelled with its own directive name. On self-hosted it also reads back what the edge is actually serving next to what you saved, flags any drift in colour, and offers a one-click adopt, so a hand-edited vhost or a value the sanitizer dropped can't quietly disagree with your settings. Self-hosted only.

The containerized edge shipped without the MaxMind geo library or its country database, so every lookup returned nothing: no per-country analytics were recorded, and country-based route rules failed closed (a single-country ban silently blocked all traffic). The library and a GeoLite2 database — parsed and validated before shipping — are now baked into the edge image, so geo works from a box's first install with no runtime download.

Behind a proxy the edge treated the proxy as the client, which broke four things at once: visitor country resolved to the PoP, distinct-visitor counts collapsed, per-IP rate limits put the whole planet in one bucket, and IP bans banned a PoP instead of a visitor. The edge now recovers the real client address from CF-Connecting-IP, trusting only Cloudflare's published ranges as the connecting peer.

Pushing a project's route rules to a local edge now goes through the same resolved edge-management path as the rest of Openship instead of assuming the edge lives on the API host's own loopback. On setups where it doesn't — such as the containerized edge — rate-limit, ban, and country rules now actually take effect.

Edge readiness is now read from whether Openship's edge container is actually running, the same fact the Infrastructure tab and System Health use. Before, a leftover bare-host OpenResty install could make the pill claim "ready" while the server tab reported the edge as down.

Openship Cloud's shared edge now forwards a free subdomain to your server only after that server answers a challenge proving it owns the target, a proof that lasts 90 days and that Cloud re-checks about a week before it lapses. Openship prepares the box to answer the moment you set up its edge, so a free domain added later works without a redeploy, and a background sweep re-asserts, reads back over the server's own public address, and re-verifies anything expired — so a box left undeployed for months no longer silently loses its free URLs. A slug can't be pointed at someone else's box, and a server that can't yet answer is told why. Requires Openship Cloud.

Adding a free .opsh.io domain to a static project used to register nothing on the edge, so the URL fell through to the wildcard with no origin behind it; editing a static project's domain wrote no vhost until the next redeploy. Both now emit a real route the instant you save, serving the project's built files — the same behavior proxied apps already had.

If an ACME order completed but a later step failed (the vhost rewrite, the reload, or the read-back), the domain stayed stuck in "provisioning" with no expiry recorded, which made it invisible to the renewal sweep — a valid certificate then sat on the edge and expired ~90 days later. Openship now re-reads the edge after a failed issuance and records a certificate that's actually present, so renewal stays scheduled.

Setting a custom domain to something that isn't a public hostname (localhost, a bare IP, a single label, or a value with a path/port/scheme) used to return success, mint a dead pending-domain row, and even write a server_name localhost; vhost onto the shared edge. The write paths now validate custom hostnames and refuse bad ones, while still letting a project that already holds a bad value edit and remove it.

A single edit could sync the same free subdomain to Cloud twice, and the two racing challenges reset each other's token, so the route worked but the project reported "Action Required" anyway. Callers that run their own managed-edge sync now suppress the duplicate, so each edit issues one challenge and the status reflects reality.

Raising a route's upload limit or adjusting its timeouts now applies as soon as you save instead of waiting for the next redeploy, and a domain redirect is applied live to static-site domains too, not only to proxied apps.

Setting up mail no longer installs iRedMail's daemons directly onto the host or reboots the box. Postfix, Dovecot, spam and virus filtering, and DKIM now run inside a single openship-mail container with its vmail database in a Postgres sidecar, host-networked so mail still sees real client IPs and the edge stays out of the mail data path. Setup is down to eight steps whose only prerequisites are Docker and the openship edge, and every piece of mail state (queue, maildirs, DKIM keys, config) lives on host bind mounts, so upgrading is a pull-and-recreate that keeps your data. Self-hosted only.

Mail setup no longer stands up its own host OpenResty and certbot on the side. The certificate for mail.<domain> is issued through the shared openship edge, the same path every other domain's TLS takes, and it is now recorded for renewal like any other domain — a mail certificate on a containerized edge (no host certbot timer) used to simply expire ~90 days after install, silently; a failed renewal now raises a notification.

The move to containers is topology-aware, so a server whose mail was installed the old way (system daemons, database on the host) still reports accurate component health and stays fully manageable: SQL queries, outbound SES/SMTP relay config, and daemon restarts all target the right engine. When the engine genuinely is stopped or was never installed, admin actions now return a clear message telling you to restart it or re-run setup instead of a bare 500 leaking a Docker "No such container" error.

A new Infrastructure view lists each server's edge and mail container with the image version it's running against the version this release pins (for example edge 0.4.0 → 0.5.0), and flags which are behind, stopped, or missing. An attention rollup surfaces the ones that genuinely need hands — an edge that's down, or absent on a box that hosts projects — separately from the ones that merely have an update available. Self-hosted only.

Update swaps a behind container onto the pinned image (rollback-guarded), Fix restarts a stopped one in place, and a server with no edge gets a one-click Install that runs the first- install path with 80/443 takeover consent. Each apply streams its pull/recreate/ verify progress and survives a dropped connection, and a fleet-wide Update all / Restart stopped does the same across every server at once. Self-hosted only.

A new instance-wide toggle lets the control plane update every remote edge and mail container automatically when you upgrade Openship (its APP_VERSION moves forward), so your fleet's infra tracks the version you're running with no manual sweep. It's server-side, so it works on desktop too, distinct from the desktop app's own auto-update, and off by default. Self-hosted only.

With auto-scan on (the default), opening the Infrastructure tab or the home page quietly runs one detect-only scan when the cached state is stale, throttled to about every 30 minutes per browser, so the attention dot reflects reality without pressing Scan. It only refreshes what's shown — it never applies an update, which stays the separate auto-update toggle. Self-hosted only.

A server converted to the containerized edge kept listing its leftover host OpenResty as a component with an Update button for a config nothing loads; those superseded host modules are now dropped from the view. And a legacy host-native mail engine (systemd Postfix/Dovecot) that was quietly delivering mail used to show up as "Stopped, container missing" — it's now labelled as not containerized rather than inventing a version. Self-hosted only.

When you move or take over a Docker stack, the reverse-proxy tunables the source vhost declared (upload size limit, upstream read timeouts) are now carried onto the migrated project before the new edge vhost is rendered. Previously the site silently reverted to nginx's 1 MB / 60 s defaults at cutover, so the first large upload after a move failed with a 413 and nothing tied it back to the migration. Adoption is additive: a limit you set by hand always outranks a value inferred from the old box. Self-hosted only.

A compose service that published a container port on a random host port (a single-part ports: entry, e.g. a database's 5432 the source never truly exposed) is now left unexposed during adoption instead of re-published on the target, so a migration doesn't needlessly expose internal services or collide with whatever already holds a host port. The port is still shown in the wizard so you can route to it, and service-to-service traffic keeps resolving by name over the shared network. Self-hosted only.

When adopting a foreign stack, environment variables that exactly match the image's baked-in defaults are dropped rather than imported as explicit config (the image still supplies them at runtime). The wizard now lists exactly which keys were skipped and why — common with Coolify/Nixpacks images that bake config as ENV layers — so you know what to re-enter if you later change the image.

The wizard used to invent a hostname from the project name whenever a step had no route chosen, labelling port fields, previewing monorepo sub-app hosts, and even wiring the "Visit Site" button to a host that would never exist. It now shows only hosts your config actually names, and hides "Visit Site" entirely when the deployment has no public domain.

Every audit entry now records where the action came in from (the dashboard, an MCP client, the CLI, the API, a webhook, or the system itself), and you can filter the log down to a single source. Isolating MCP-driven activity finally answers "what did the AI assistant do," which was impossible before because an MCP write and a CLI write produced identical rows. Entries written before this release show as an honest "unknown" rather than being mislabelled.

Each organization can switch audit recording on or off and pick a retention window (7, 30, 90, 180 or 365 days). Turning recording off stops new entries but never touches the ones already written, and the act of disabling is itself logged first, so the trail always explains why it went quiet. The switch is per-organization, so on a shared cloud instance one tenant can't disable everyone's audit log.

You can narrow by category, actor, date range and free-text search, with a live count on every category tab. Search resolves project, server and domain names, so typing a name like "api-gateway" turns up entries that only ever stored an opaque id, and rows now display the human name of the affected resource, resolved server-side in one batched lookup per page.

Entries that share a timestamp are now ordered by a stable tiebreaker, so scrolling through pages no longer reshuffles them; two entries created in the same instant could previously appear twice or vanish between pages.

Telegram joins email, Slack, Discord, and Teams as a notification channel: point a BotFather bot at a chat, group, or forum topic and Openship delivers your deploy, health, and job alerts there. The bot token is stored encrypted, and the channel list shows which bot does the sending.

On a self-hosted box the billing notification categories are fed by the SaaS billing system, so they were switches that could never send anything. They no longer appear, and the remaining categories are organized into tabbed groups in Settings.

The curl | sh and PowerShell installers download a self-contained, sha256-verified CLI payload that runs under Node: they use your system Node 22+ if you have one and otherwise vendor an official Node 22 from nodejs.org into ~/.openship, and they quietly migrate you off any previous global-Bun install without touching your own Bun. openship update takes the same path — re-download the verified payload, refresh the runtime first — so an upgrade that raises the Node floor heals itself instead of breaking. This retires the Bun-vs-ssh2 crash class that could take out remote SSH operations.

Openship's official multi-arch (amd64 + arm64) image set, built and published to GHCR and Docker Hub each release, now includes openship-mail alongside the api, dashboard and edge images, so a self-hosted mail engine no longer has to be built from source.

Bringing up the compose stack on a rootless daemon used to abort the install with Docker's opaque "error while creating mount source path … permission denied," because a rootless daemon can't create the edge's bind-mount directories under root-owned paths. The CLI now creates those directories itself and, when it can't, prints the exact one-time sudo mkdir/chown commands to run (#372).

Build output that carried a raw NUL byte or a broken Unicode character (as a failed docker exec does when it spills its multiplexed stream, frame headers and all) made the log write fail, and Openship misread that rejected write as the deploy itself failing: a live, running deploy was marked failed and its containers torn down, while the deploy view stayed stuck on "Deploying." Stored build logs are now sanitized and size-capped, the outcome write sheds any log or error blob Postgres refuses rather than losing the status, and the terminal event that closes the live stream always fires.

A project's deploy target is now stored durably on the project itself (a new server_id column) instead of living only in the latest deployment's volatile metadata. Before, a fresh or partial snapshot could lose that binding, quietly regress the project to a local deploy, strip the ports off its custom domains (a 502), and reset the Access URL to localhost:3000; existing server-hosted projects are backfilled so this can't bite them on the next redeploy. A verified custom domain is also preserved rather than deleted or blanked even when a deploy's endpoint set omits it. Self-hosted only.

A box set up with a free .opsh.io domain could finish onboarding with no servers listed and nothing to deploy to, because registering the local "This Server" row hung off a bootstrap step that path never runs. The box now materializes its own deploy-target row when servers are first listed, when it connects to Openship Cloud, and on admin reset. Attaching a free .opsh.io domain to a project later no longer needs a redeploy first, since every deploy now primes the box to answer the ownership check. Self-hosted only.

A compose service's own domain ignored the project's reverse-proxy tuning, so the main app would accept a 50 MB upload while the same project's service domain rejected it at nginx's 1 MB default with a 413. Those settings now apply to every domain a project writes — per-service, static, single-domain composites, path fan-out, and single-app native deploys. Self-hosted only.

Apps that route one container port to a domain and leave another port-only (Convex, for one) used to get an empty or self-referential origin injected (CONVEX_SITE_ORIGIN="", or a 127.0.0.1 that inside a container points at the container itself). Each port now resolves independently, a port-only service falls back to the box's real reachable http://host:port, and a {{publicUrl:…}} token that still can't resolve is left unset — so the image's own default applies — with a loud warning in the deploy log instead of a silent blank.

When a deploy succeeds but its domains aren't being served, the post-deploy warning now checks why: if the edge container itself is down it tells you to bring the edge back up, instead of sending you off to your DNS provider to debug routes that are actually fine. Self-hosted only.

The Rollback action used to be greyed out on projects using the default settings, because nothing marked their releases as restorable. It's now available on every successful release, and it is a single action: Openship reuses the release's retained image when it's still on the server (seconds, no rebuild) and rebuilds that release's commit when it isn't. It can no longer dead-end — the "Redeploy this commit" fallback button is gone because the one action already covers it, and the confirmation tells you which of the two you're about to get.

A rollback now runs the same deploy pipeline a normal deploy does, replaying that release's own frozen configuration and environment variables. Previously it hand-assembled a bare container, which meant a "successful" rollback could come back with no environment variables, no published port (a 502 behind your domain) and its volumes detached. Restores now get the health check, the crash-loop watch, routing, logs and a build log of their own, exactly like any other deploy.

Only services whose image is missing are rebuilt; a service the deploy never touched keeps running on the image it already has, so rolling back your app code doesn't bounce your database.

A static release's artifact is its built files on disk, not an image, so restoring one copies that version's files back into place (hard-linked, so it costs almost no extra disk) and re-points the edge at them. No image, no rebuild. A static rollback previously tried — and failed — to restart a directory as if it were a process.

A restore reuses its source release's image, so two releases can share one. Retention now knows that and will not delete an image another retained release still needs.

How many releases stay restorable is now measured from free space on the deploy host and your project's real image size (a quarter of what's free, between 2 and 20 releases), instead of a fixed 5. You can still pin an exact number; clearing the field returns it to automatic.

Rollback retention now lives in the project's Backup tab, and in the deploy wizard's target panel where you pick the server, with the measured snapshot size and free disk shown inline. Pinning a release to keep it restorable indefinitely is unchanged.

The wizard's rollback controls are editable on a first deploy too, and the choice is applied when the project is created. They used to render read-only until the project existed, which was the one moment you were actually looking at them. The card also names what a retained version is on your project — built files for a static site, images otherwise — instead of talking about images either way.

It listed only the build location while hiding the rollback window and clone location; it now names each section it contains. The summary chips next to the target (Static, Sandboxed, tier) lost their coloured pill backgrounds and read as plain text, with the one that matters — an unsandboxed "direct on host" runtime — still called out in colour.

It used to be frozen onto each deployment as it was created, so turning on instant rollback did nothing for anything already deployed.

A release whose configuration snapshot predated the "production paths" setting crashed the deploy pipeline outright. This affected plain redeploys too, not just rollback.

A release that reuses an existing image needs no install or start command, but pre-deploy checks demanded them. Adding a required setting would otherwise have made every older release un-restorable.

A stack adopted from a Docker migration has no git repository and needs none, but deploys were refused for not having one.

Six of the nine places that write per-service deployment records left the service name blank, so a later rollback couldn't tell services apart and rebuilt the whole stack.

An empty value now falls back to the default instead of purging every restorable release.

Local deploys honor DOCKER_HOST (and an explicit socket path) instead of assuming /var/run/docker.sock, so Colima, Rancher Desktop, Podman and rootless Docker work.

The MCP server now ships a prompt catalog (deploy-from-git, deploy-a-folder, install-catalog-app, and an orientation overview) so an AI client follows the correct tool sequence instead of reverse-engineering it from a flat list. Its write tools now carry typed request bodies end to end (projects, deployments, domains, webhooks, notifications, connections, apps), and every prompt points at github.com/oblien/openship/issues when a client hits a platform bug.

A token granted a single GitHub repo (or the "create your own projects" scope) now correctly advertises the tools it can actually call. Previously the per-repo GitHub tools and the project create/list tools were filtered out of tools/list, so a scoped token saw nothing to work with.

"Include www" always created a second hostname, but the pieces around it still treated the pair as one thing. Renewing SSL for a domain issued the www certificate inside the same operation, unguarded: a www that wasn't pointed at the server yet failed after the apex had already succeeded, and the apex was reported as broken. Adding a domain with the switch on also showed you only the apex's DNS record, so www never resolved, its certificate could never be issued, and every deploy retried a hostname that had been set up to fail. Both hostnames now get their own DNS record, their own Verify button, their own certificate — and their own failure.

Any domain can now answer a 301 (or 302) to another domain of the same project, set on the domain's card. www uses it by default (www.example.com → example.com), and the direction is yours to flip or turn off. Old-domain-to-new-domain moves work the same way. The path and query string are preserved, the redirecting host still gets its own certificate, and a target outside the project — or a redirect that would loop — is refused.

The verify modal streamed certbot's output and then, on any failure, replaced the whole console with one line: "the connection closed before the operation reported a result — check the domain's status." The actual reason was discarded. The log now stays on screen in every outcome, with Copy log and Try again next to it. And if the connection really does drop, Openship reads the domain's state back and tells you what happened instead of asking you to go and look.

A keep-alive ping could race the final event of any live-log stream (verify, edge setup, deploys) and win, so the browser never received the result of an operation the server had already completed. Stream writes are serialized now, and a result that has arrived can no longer be overwritten by the connection closing behind it.

Certificate issuance is serialized per server, not just per hostname. www made two pending domains the normal case, and a manual Verify could collide with the background retry working on the sibling: both ran certbot, both wanted the same challenge port, and one died with an error that read like a DNS problem.

On a self-hosted install the pre-deploy DNS panel now fills the A record's value with the server's public address (detected once when the server is registered) instead of leaving it blank.

The panel no longer tells you to add a TXT record on a self-hosted box; there isn't one — HTTPS is issued automatically on the first deploy. The DNS modal is also lighter and on-theme, and the "Include www" switch now sits below the domain field so it no longer shifts the input as you type.

Bare openship (and the from-source openship-dev) now recognizes a Docker Compose install (the default on Linux). Re-running after setup manages the running stack instead of walking you through name / email / domain from scratch again.

On a compose install these now run docker compose up -d / restart / down and read the stack's real state, instead of targeting a systemd unit that was never installed (which reported "stopped" for a healthy stack).

Migrating from another platform (Coolify/Dokploy/Dokku) or adopting a running Docker host could stall the reachability check — or drop the request outright — under the Bun-hosted API, even against a perfectly healthy daemon. The bridge now starts reading the request socket immediately and verifies each forwarded channel actually carries data, falling back to docker system dial-stdio on a fresh connection when a channel opens dead. Contributed by @jbermudez00 (#271).

The iRedMail engine is now shipped inside the packaged desktop app (and the CLI bundle) and located by an explicit path, fixing the Transfer iRedMail Engine … tar: could not chdir failure on install.

A remote (VPS) instance pastes an access token inline in the Library, with no gh auth login hints; the gh CLI path is now desktop-only, where it belongs.

The deploy-target picker no longer flashes on entry. A sensible target is applied silently and stays one click away in the summary bar at the top.

The self-managed "Openship" project no longer shows a bogus public openship service (and its stray openship-openship.opsh.io route) that matched no container and read "Stopped" forever.

The edge now owns a 443 default server that refuses any hostname it doesn't route (one you removed, never added, or merely pointed at the box's IP). Before this, such a request fell through to the first-loaded vhost and was served another app's certificate and backend. Applies automatically on the next deploy, on both the bare and containerized edge. Critical — see the in-app advisory.

A migrated compose stack redeploys like any repo project: it reclones and rebuilds build: services and pulls image: ones, instead of failing on a frozen build tag (404 no such image). The running image is reused only once, at cutover.

The migrate screen lists every repo compose service, not just running containers, so a service with no container (e.g. redis, or an app that wasn't up) is built/pulled and routed like the rest, with its env and route editable on the card.

A reused container is joined to the new project's network under its service-name alias, so a freshly-built app resolves postgres:5432 by name, exactly like a native deploy.

Service state is read live from the host and matched by identity (label → openship-<slug>-<svc> name → tracked id → compose labels), so a container Openship adopted in place (its docker labels still name the previous project) no longer shows "Stopped" while it serves traffic. Each run's log now ends with the container, state and match for every service.

Start/Stop/Restart, logs, terminal, backup/restore and volume sizes resolve the container against the host first, so a redeploy that replaced it no longer leaves them failing with no such container — or, on Start, launching a duplicate container beside the running one. A crash-looping container now reads Restarting instead of a green "Running", and an unreachable host reads Unknown instead of echoing the last deploy status.

The free-domain gate classifies by hostname, so removing a custom-domain route (or any route) is no longer blocked by an unrelated free subdomain still in the set.

Runtime teardown is time-bounded, so a slow or unreachable server no longer strands the delete before the record is removed.

MacOS now ships separate arm64 and Intel x64 dmgs (both built and SHA-256-checksummed in CI), so Apple Silicon Macs run natively instead of under Rosetta. Windows (x64) and Linux (AppImage) are unchanged.

The stack detector now reads railway.toml/railway.json and vercel.json (build / install / start / output commands, framework, and routing) and folds them over its own detection. A repo that already tells Railway or Vercel how to build it deploys the same way here, no reconfiguration. Every config source runs through one shared parser registry (no per-source special-casing).

An optional repo-root config to declare build, routing, env, and domains up front; it's authoritative over auto-detection and rides the same engine, for the repo root and each monorepo sub-app.

A self-hosted instance targets the server it runs on, never Openship Cloud, unless you explicitly choose cloud.

The post-deploy probe reaches your app through the host gateway, so a containerized self-host no longer fail-reverts an otherwise-healthy deploy.

The edge install no longer pins the APT repo to a codename that doesn't exist yet (e.g. Ubuntu 26.04), and self-heals a box already broken by the old pin.

The service and its children are reaped by process group and any ports it held are swept, so a restart can't strand the old process on a new port.

The Pages output path resolves through one shared, sandboxed resolver so a build can't escape its output dir.

A fixture only, never a shipped dependency.

Installed apps now expose a curated settings surface (schema-driven) so you can change an app's real config after install without digging through raw env. Edits go through a safe env-merge and tell you whether a full redeploy (vs a quick restart-apply) is needed.

Clicking a catalog app opens a focused, business-only setup that creates the project on confirm; the technical deploy wizard is now the "Advanced" path (no more orphaned draft projects from a half-finished install).

It appears in the catalog alongside Convex and n8n and hands off to the mail wizard. The rest of the catalog shows as Coming soon (dimmed, not installable) for this release.

(Read-only) — backup schedules run on the same job runner as everything else (zero duplication), so their next/last run sits right next to your system and custom jobs.

Each self-hosted server now authenticates to GitHub on its own, from a dedicated GitHub tab: sign in with a device code (like gh), paste a token, generate an SSH key to add to your account, or use auto-registered read-only per-repo deploy keys. Credentials are stored encrypted and the exact same connect panel is reused inside the deploy flow, so a missing credential is one click to fix mid-deploy. Private-repo clones now work without your desktop online.

Per-destination storage stats, a sticky status rail, and clickable rows that open a per-destination detail page showing exactly which projects and services back up there.

Openship Cloud enforces a hard cap on projects per user (env CLOUD_MAX_PROJECTS_PER_USER, default 2), at both create and folder-upload/ensure. Self-hosted is unmetered.

Project deletion shows a real Deleting state, and when the source teardown can't complete you get a clean "Delete from storage" option that drops the record immediately (leftover resources are reclaimed later by GC). The atomic, all-or-nothing delete stays the default.

The login redirect now lands on the same loopback host the session cookie was minted on (localhost ⇄ 127.0.0.1), so the dashboard no longer opens cookieless and bounces you back to /login.

No more false "locked by a different machine" on your own box when the machine-id probe is momentarily flaky; the cross-machine guard now only fires on a genuinely different, stable machine id.

Status colors (success / warning / danger / info) are unified semantic tokens across the whole dashboard, and the dim theme's greens and reds are tuned for comfortable contrast.

Clearer illustration, a See docs action, and a distinct icon per "what gets configured" tile.

Unified page-style header with the Continue action aligned to the config column, and a collapsed, searchable server picker (with an inline "Add your own server").

Pnpm/yarn are now enabled via corepack across every build path (cloud, generated Dockerfile, bare host, monorepo workspace-prepare, cloud local-build). Fixes pnpm: not found on deploy.

, Expanded to 15 production-ready self-hosted apps: Convex, n8n, Ghost, Directus, NocoDB, Metabase, Grafana, Gitea, code-server, Uptime Kuma, Vaultwarden, FreshRSS, Stirling PDF, IT-Tools, Excalidraw.

A verifiable pending row is minted on add/create/edit, one canonical hostname normalizer is shared across storage/routing/domain-service, lookups are cross-tenant-safe, and certbot is gated on verification (no wasted Let's Encrypt attempts).

Connect-on-server credentials honored in preflight.

Is only offered where it works (team orgs on a multi-user instance); single-user/personal instances are guided to migrate or create a team org instead of hitting a dead end.