The story behind Drift FM — why it exists, why it’s built the way it is.

For the technical details, see Architecture. For the frontend design system, see Design Language.

The Problem

Every music app wants your attention. Recommendations, social feeds, notifications — they optimize for engagement, not listening. I wanted something that plays music for a mood and gets out of the way.

Drift FM has one interaction: pick a mood. After that, music plays. No decisions, no feed, no algorithm nudging you toward something else.

Why Self-Hosted

Your files, your server. No subscription, no licensing gaps, no “this track is no longer available.” No analytics, no cookies, no third-party scripts.

It runs on a $5 VPS, a Raspberry Pi, or localhost. The binary is ~15 MB.

The Stack

Go + SQLite + vanilla JS. Boring on purpose.

Go compiles to a single binary. Cross-compile to any platform. The stdlib HTTP server is more than enough for serving files to a handful of users.

SQLite because a few thousand tracks is a tiny dataset. WAL mode for concurrent reads. One file to back up. No Postgres, no connection pooling, no Docker Compose. We use modernc.org/sqlite — pure Go, no CGO, clean cross-compilation.

Vanilla JS because the player is one page with a few interactive elements. A framework would add a build pipeline and a package manager to solve problems this app doesn’t have. CSS variables handle theming. The <audio> element handles playback.

Shuffle with Memory

True randomness feels repetitive. Hearing the same track twice in an hour feels broken, not random.

The shuffle uses recency avoidance:

  1. Partition tracks into “not recently played” and “recently played” (last 3)
  2. Fisher-Yates shuffle only the non-recent tracks
  3. Append recent tracks at the end, unshuffled
  4. When a track plays, add it to the recent list (FIFO, capped at 3)

Simple, works well for small libraries. Stateful per mood — switching moods resets the window.

The .txt Convention

Focus mood enforces instrumental-only. Rather than import flags, we use a file convention: if a .txt file exists next to an .mp3 with the same name, the track is vocal. Content becomes displayable lyrics. Empty file still marks it as vocal.

Zero-config. Drop files, batch import, the script figures it out. See Quickstart — Vocals and lyrics for examples.

Deploy

One binary, one database file, one directory of audio files. Copy three things, run it. See the README for the full deploy guide.