October in one paragraph

October’s changelog has 30 entries, dated October 1 through October 31. I grouped them into five themes: isolated environments, scaling and paying for what runs, safer deploys, seeing what is happening, and agents and code. The common thread is that what runs in production stays isolated, observable, and predictable. The table below links every entry. The full changelog is the complete history, and company is separate from it.

Isolated environments and previews

A preview stays off production only when its variables, its network, and its lifetime are separate. How preview environments stay isolated is the longer note. October added the three pieces below.

Environment-scoped variables

Every service or project variable is All, Production, or Preview. A scoped value overrides All for the same key, so a preview can point at a staging database or a test API key. Existing variables stay All. Closing a pull request tears that preview’s stack down. Environment variables for production and previews.

Switch, copy and promote environments

Every project has production, custom, and preview environments. You can switch, copy any environment, or start an empty one. Pull request previews are preview environments of the project, not separate projects. A self-hosted database volume starts at 1 GB. Switch, copy and promote environments.

Private services, and previews can’t reach production

A private service has no public URL, domain, or certificate. Other services in the same environment reach it on its own port. Traffic stays in the same organization and the same environment, so a preview cannot reach production, including another project’s production. Private services and isolated environments.

Scaling and paying for what runs

Replicas are billed while they run, Free apps can sleep, and a failed card payment does not stop services the same day. How capacity-based billing works and what Free and Pro include.

Autoscale on CPU, memory or both

Autoscale sets a minimum and a maximum number of replicas. Scale on CPU, on memory, or on both, each as a percentage of the per-replica limit. When both are on, the one asking for more replicas wins. You are billed for the replicas that are running, not for the maximum. Autoscale on CPU, memory, or both. How to set the minimum, the maximum, and the trigger is in Replica autoscaling.

On-demand apps that sleep and wake

Free apps sleep after a period without traffic or active requests, and the next request wakes them without a new deployment. The entry does not say how long that idle period is. They no longer pause after 7 days of inactivity. Pro services stay always-on unless you turn On-demand on for that service. Sleeping time does not count toward compute usage. On-demand apps that sleep and wake.

A grace period for failed payments

A failed card payment leaves services running for three days. The charge is retried on day one, a final warning goes out on day two, and running services pause on day three if the balance is still unpaid. Nothing is deleted. New resources cannot be created while the payment is outstanding. A grace period for failed payments.

Safer deploys

A push should wait for its checks, skip untouched paths, and stop if the new version cannot start.

Wait for CI

Wait for CI deploys a commit only after every GitHub check on it has passed. A commit with no check is deployed after two minutes. If CI never finishes, the commit is dropped after two hours. Wait for CI before deploying.

Watch paths

Watch paths choose which files trigger an automatic deploy, independent of the build root. An empty list keeps the previous behavior. Manual deploys, deploy hooks, and pull request previews ignore watch paths. Watch paths for automatic deployments.

Crash-loop protection and slow first boots

A service that keeps restarting with no healthy replica can be paused automatically, even when health alerts are off. A new version gets up to 10 minutes to become ready. A version that cannot start fails right away, with the reason. Crash-loop protection and October fixes.

Post-deploy commands

On a paid plan, a repository service can run one command after the new version is serving traffic, for up to 10 minutes. A failure leaves the release up and rolls nothing back. Image services cannot use it. Run a command after each deploy.

Build plans, Railpack and clearer build failures

Each deployment records a build plan before the build starts, including secret names and never the values. The builder order is a configured Dockerfile, then railpack.json, then Unbroker’s generated build, then Nixpacks. A failed build comes with a reason and a next step. Build plans, Railpack and clearer build failures.

Seeing what’s happening

October added searchable logs, alerts, a Postgres data browser, pgvector on new template databases, and project traffic. Postgres in production is the separate note on connections and restore.

Log search, alerts and Postgres Insights

The Resources tab ranges from 15 minutes to 30 days. Logs can be searched and filtered by level. Health alerts go to the project’s email and Slack channels, and the Health alerts switch requires Pro. Self-hosted Postgres gets an Insights tab. Log search, health alerts and Postgres insights.

Readable logs

Build logs show one row per step, with its duration, or cached when the step was skipped. JSON runtime lines are unpacked into a message, the logger name, and the other fields. Build and runtime logs you can actually read.

Traffic analytics

Analytics shows visitors, page views, requests, and sources for a project, then a production or preview environment, then a service. It requires a paid plan and a cell with traffic collection enabled. It reads HTTP requests at the ingress, with no browser script. Traffic analytics for each project.

Postgres data browser, pgvector and Postgres 18

The Data tab on self-hosted Postgres is read-only. New databases from the template use the Postgres 16 image with pgvector. Run CREATE EXTENSION vector; to store embeddings. Existing databases keep their image. Postgres 18 or newer, with the volume at /var/lib/postgresql/data and no PGDATA of your own, gets PGDATA set. Versions 17 and older are not changed. PostGIS, TimescaleDB, and pgvector images are recognized as Postgres. Browse your PostgreSQL data and pgvector on new Postgres.

Agents and code

You can connect a coding client to the organization, or describe the environment in a file and apply it after a plan. Letting a coding agent deploy safely and Unbroker for coding agents.

MCP server and live dashboard

The remote MCP endpoint is https://api.unbroker.cloud/mcp. Owners and admins turn it on per organization, and it uses your existing API tokens. Guides cover Cursor, VS Code, Claude Code, Codex, and Claude Connectors. Inspection tools do not return credentials or environment variable values. The dashboard updates without a reload. MCP server and a live dashboard.

Infrastructure as code with ubctl

An environment can live in .unbroker/, in TypeScript, Go, or Python. ubctl plans, then applies. Pull does not export secret values. Apply asks before anything destructive and deploys only services whose runtime changed. Infrastructure as code.

Deploy on Unbroker button

A Deploy on Unbroker button in a README opens a ready-to-deploy form for a catalog template or a public GitHub repository. Someone signed out signs in and comes back to that form. Private repositories still go through New project. Deploy on Unbroker button.

Audited commands

Manual command and job runs record who launched them and which service they belong to. The command text stays on the run, not in the organization-wide activity feed. The interactive browser terminal is gone. One-off commands and scheduled jobs remain. Audited commands and a simpler Jobs tab.

Also new: the blog, and the site in Spanish and Portuguese

The website is available in English, Spanish, and Brazilian Portuguese. The footer switcher covers the home page, pricing, guides, and comparisons, and prices come from the same catalogue in every language. The blog publishes in all three languages, with an RSS feed for each. A blog, and the website in Spanish and Portuguese.

Full list

Every October entry is one row. Two entries that share a date are both listed.

DateChangeEntry
2026-10-01A grace period for failed paymentsOpen
2026-10-01Browse your PostgreSQL data from the dashboardOpen
2026-10-02Log search, health alerts and Postgres insightsOpen
2026-10-02Steadier deploys and metricsOpen
2026-10-05Build and runtime logs you can actually readOpen
2026-10-05Wait for CI before deployingOpen
2026-10-06Autoscale on CPU, memory, or bothOpen
2026-10-06Environment variables for production and previewsOpen
2026-10-07Choose your service’s addressOpen
2026-10-07A warmer welcome and a handier user menuOpen
2026-10-08Audited commands and a simpler Jobs tabOpen
2026-10-08Manage Google and GitHub sign-in from your profileOpen
2026-10-10Crash-loop protection and deployment fixesOpen
2026-10-12Traffic analytics for each projectOpen
2026-10-13Product guides for choosing infrastructureOpen
2026-10-14Switch, copy and promote environmentsOpen
2026-10-15Deploy on Unbroker buttonOpen
2026-10-15MCP server and a live dashboardOpen
2026-10-21Watch paths for automatic deploymentsOpen
2026-10-22Infrastructure as codeOpen
2026-10-23Groups on the board and a faster projects pageOpen
2026-10-24pgvector on new Postgres, plus Postgres 18 imagesOpen
2026-10-26A blog, and the website in Spanish and PortugueseOpen
2026-10-26Build plans, Railpack and clearer build failuresOpen
2026-10-26October fixes and improvementsOpen
2026-10-26On-demand apps that sleep and wakeOpen
2026-10-26Run a command after each deployOpen
2026-10-26Private services and isolated environmentsOpen
2026-10-27More resilient Git buildsOpen
2026-10-31Static framework sites on FreeOpen

FAQ

What did Unbroker ship in October 2026?

30 changelog entries, dated October 1 through October 31. This page groups them into five themes and links each one. The full changelog is the complete history.

Does Unbroker support autoscaling?

Yes. Autoscale sets a minimum and a maximum number of replicas and scales on CPU, memory, or both. You are billed for the replicas that are running, not for the maximum. Your plan sets how far a service can scale. See how capacity-based billing works.

Can I keep a service private on Unbroker?

Yes. A private service has no public URL, domain, or certificate. Other services in the same environment can reach it. Traffic stays in the same organization and the same environment, so a preview cannot reach production.

Can I use different environment variables for production and previews on Unbroker?

Yes. Each variable is All, Production, or Preview. A value scoped to an environment overrides All for the same key. Existing variables stay All until you change them. See how preview environments stay isolated.

Does Unbroker have an MCP server for Cursor, Claude Code or Codex?

Yes. The remote endpoint is https://api.unbroker.cloud/mcp. Owners and admins turn it on per organization, and it uses your API tokens. Connection guides cover Cursor, VS Code, Claude Code, Codex, and Claude Connectors. See Unbroker for coding agents.

Can I define Unbroker infrastructure as code?

Yes. A project environment can live in .unbroker/, written in TypeScript, Go, or Python. ubctl shows a plan, then applies it. Pull does not export secret values.

Do free apps on Unbroker sleep?

Yes. Free apps sleep after a period without traffic or active requests and wake on the next request. The entry does not say how long that period is. They no longer pause after 7 days of inactivity. Pro services stay always-on unless you turn On-demand on for that service. See what Free and Pro include.

Where can I see every Unbroker change?

The table on this page lists every October entry. The full changelog lists every release, including months other than October.