Ali Husain Software engineer and product designer

03 Infrastructure/Vesper

Vesper

Self-hosted app · open source

A read-only operations dashboard that answers one question: does anything need you?

14 seconds of the dashboard on demo data: one service goes down.

Vesper on a phone when something needs attention: one file differs from the repository.Vesper on a phone when something needs attention: one file differs from the repository.
Vesper on a phone: nothing needs you, 12 of 12 services up.Vesper on a phone: nothing needs you, 12 of 12 services up.

One Go binary reads what the platform's scripts already record and says it in one sentence. The ribbon is the fleet drawn as light, one lane for each service, so an outage can be seen turning. When Vesper cannot confirm something it says so instead of showing an all-clear, and it is read-only by construction.

7.4 MB
one binary, with the interface inside
27 MB
of memory on a Raspberry Pi 5
191
automated tests, 34 of them end to end under real systemd
0
Go dependencies beyond the standard library

Why it exists

A home server run with care records a lot about itself. The backup job writes what it verified, a weekly restore drill (a test restore of the archives) proves it, and a notifier remembers every alert. These facts sit in separate files and logs. No one place says whether anything needs me.

What it does

Vesper is a read-only operations dashboard for a home server. It reads those sources, probes each service as a client would, and opens with one sentence: "Nothing needs you", with the checks behind it, or the problem by name, with its checklist.

If it cannot read a source, it says "Vesper can't see". It gives no all-clear it cannot prove.

How it is built

  • Two processes, one snapshot

    The collector runs as root with only the Linux permission to read any file, and writes a snapshot to memory every 20 seconds. The web process has no network and no capabilities, so a flaw in what a browser reaches is contained, away from the server's files and network.

  • Unknown is its own state

    A five-minute slot with no sample stays unknown, so uptime counts only watched time. A check that is set up but never ran is a blind spot, not a pass.

  • Standard library, one binary

    The Go code has no dependencies and embeds the interface: one binary to ship, no database to run.

  • The demo runs the real generator

    The live demo is the code behind vesper demo, compiled to WebAssembly. CI checks that both return the same snapshot.

How I know it works

191 tests
60 Go tests under the race detector, 97 interface tests and 34 end-to-end checks under real systemd, on every change.
7.4 MB
The static arm64 binary on the releases page, interface included.
0.5
The web process's exposure score from systemd-analyze security, where lower is safer. The collector scores 2.4.

What was hard

A back-off that had already lapsed

The password hash is slow on purpose: PBKDF2 at 600,000 rounds. After five wrong guesses a client must wait. My first version stamped a failure with the time taken before the check. Where the check takes about a second, the one-second wait had lapsed before it was recorded, and a sixth guess went straight through.

The fix reads the clock again after the check. Checks also run one at a time, so a burst of parallel logins cannot outrun the limiter. A test sends twelve wrong logins at once; at most five may be checked.

Limits

  • It cannot restart a service or run a backup, on purpose.
  • It samples every 20 seconds by default and keeps one day of history per service, so a shorter outage can go unseen and trends need a metrics tool.
  • The backup, drill, alert and drift views expect the file layout of one particular homelab, unless their paths are set in the configuration.

Also in Infrastructure

Open to remote work.

ali@gallowforge.dev