# Building several sites at once

If one corpus publishes several sites, **Build all sites** builds them **in
parallel**, in isolated containers, and then deploys them serially, whenever a
container engine is available on the machine. With none, it builds them one after
another instead. A single site's own build always runs one at a time. There is
nothing to switch on.

## What you need

- A container engine: Docker, or podman. Rootless podman is a good fit, since it
  maps container files to your host user automatically.
- The host still needs Node and pnpm — two of the three phases run there — plus
  your hosting credentials in the environment.

The build image is created and cached automatically the first time you run it.

## How a parallel build runs

One job runs three ordered phases:

1. **Shared work, on the host, once.** Build the search index and warm the
   archive cache for the union of every site's channels. Only the host writes
   this shared state, so the containers can never race each other for it. This
   phase is the long pole on a cold build and nearly instant on a warm one,
   because unchanged channels are skipped.
2. **Per-site builds, in parallel containers.** Each site composes and builds in
   its own container, up to a configured concurrency limit, writing to its own
   output directory. Containers mount the corpus and index **read-only**.
3. **Deploy, on the host, serially.** Once every build has finished, each site is
   published in turn.

## Why credentials never enter a container

Deployment runs on the host, in phase three, precisely so that hosting
credentials are never passed into a build container. A build container gets
read-only mounts, a non-root user, and its own output directory — and nothing
that could publish anything.

Network access is left enabled inside the containers, because the site build
fetches its web fonts. The isolation comes from the read-only mounts and the
separate output directories, not from cutting the network.

## When it is worth it

The parallel pipeline pays off in proportion to how many sites you publish and
how little they share. With one site it is strictly slower — you pay for image
management and gain nothing. With several, phase two collapses from the sum of
the build times to roughly the longest single one.

Next: [AI and MCP](/docs/ai-and-mcp/).