Quiver

Quiver is the host I built for the pile of small web apps that don’t justify a full Kubernetes deployment - a one-page utility, a report viewer, a 200-line toy. You drop a directory and it gets a working hostname, HTTPS, and its own access rules. The whole point is that an app you hit twice a month should cost nothing the rest of the time.

How it works

• Socket activation, zero idle RAM. A systemd .socket unit per app owns the listening socket; the app process doesn’t exist until a request arrives. A shared middleware exits the process after an idle window, and the armed socket cold-starts a fresh one on the next hit. At rest the whole box is ~35 MB; each awake app adds ~19 MB.
• An app is just a directory. A manifest picks a Python ASGI callable or any binary (a loopback proxy fronts non-cooperating binaries so they get the same activation contract), static files are served directly, and one writable data/ path holds SQLite.
• Per-app access, reconciled to the edge. Each app declares LAN-only or public- behind-SSO in its own access.yaml. A root reconciler makes the CDN match it live - DNS record, tunnel route, and an SSO application whose policy lists exactly who may sign in.
• Static artifacts and reports. It also hosts a durable static store with a searchable browser, so generated reports and one-off pages have a single findable home instead of scattering across ad-hoc servers.

Using it

From the operator side the whole appeal is that standing something up costs almost nothing in effort or in memory. You drop a directory with a small manifest, run one deploy command, and the app comes up at its own hostname with HTTPS - no repo, no Dockerfile, no CI pipeline, no route to wire by hand. Because idle apps stop and hold no memory, you never have to decide whether a tool earns its keep: a dashboard you check twice a month sits there costing nothing until the next request, so the host quietly accumulates dozens of them without you managing capacity. Access is a one-file flip - each app’s access.yaml says LAN-only or public-behind-SSO and exactly who may sign in, and a reconciler makes the edge match - so sharing a tool with one other person is an edit, not a project. Generated reports and one-off pages get the same home through a searchable artifact browser, which means the thing you made for someone to read is findable later instead of scattered across ad-hoc servers you forgot you started.

Notes

The design decision that carries everything is Restart=no plus a per-app idle-stop: an automatic restart would defeat idle-stop and pin memory forever. That single constraint is what lets the host fit dozens of apps in 1 GB.