Everything else in this category replaces a subscription. Fugleramme replaces nothing — it’s here because it’s the most enjoyable thing you can do with a Raspberry Pi and an afternoon, and because it’s a genuinely good demonstration of local AI doing something worth doing.

A mic listens to your garden. BirdNET-Go identifies the species from their calls, entirely on the device, with nothing leaving the house. Fugleramme polls its API, matches each bird to a public-domain illustration cut from a real 19th-century plate, packs them onto a textured paper page, and redraws when the birds change. The result hangs on a wall and shows the birds that are actually outside right now.

The author’s own frame is live at fugleramme.arnegiacomo.dev, showing what’s in a garden in Bergen at this moment.

You don’t need the e-ink panel

The nicest version is an Inky Impression 13.3" in an A4 frame on a Raspberry Pi 5. But the panel is optional — without one it runs web-only and the page takes the shape of whatever displays it: a TV, an HDMI monitor, a tablet on a shelf, or your desktop wallpaper. That makes it a zero-hardware thing to try first.

  docker run -d -p 8080:8080 -v fugleramme:/data \
  -e FUGLERAMME_DETECTOR_URL=http://birdnet.local:8080 \
  ghcr.io/arnegiacomo/fugleramme

Kiosk on :8080, admin on :8080/admin, state in /data. Already running BirdNET-Go? Point FUGLERAMME_DETECTOR_URL at it, anywhere on your network.

Nothing running yet — a Linux box with a USB mic brings up both:

  curl -fsSL https://raw.githubusercontent.com/arnegiacomo/fugleramme/main/examples/docker-compose.yml -o docker-compose.yml
docker compose up -d

On a Pi, the installer does the whole thing — BirdNET-Go, dependencies, and a systemd service:

  curl -fsSL https://raw.githubusercontent.com/arnegiacomo/fugleramme/main/install.sh | bash

And to poke at it on a laptop with no hardware at all, there’s a fake detector:

  uv sync
uv run fugleramme-fake-detector    # stand-in BirdNET-Go on :8090
uv run fugleramme-dev              # frame on :8080, hot reload

Why the art matters

Over 800 cut-outs covering 400+ species, every one taken from a real plate and hand-curated. None of it is AI-generated — some has been retouched, and the manifest links back to the plate each file was cut from.

The layout is doing more work than it looks: larger birds toward the centre, each sized by actual body mass (from the AVONET dataset), and an empty window drawn as a bare perch when nothing has been heard. That’s the difference between a dashboard and something you want on a wall.

Coverage follows the source plates — Scandinavian, British, and central European, so the Nordics, the British Isles, and Germany are well served and elsewhere is thinner. Wider European and North American coverage is in progress, and adding artwork is the contribution the project asks for most.

Before you build one

  • Early development. The README says so plainly: expect rough edges.
  • Licensing is layered. Code is MIT, but BirdNET-Go’s detection model is CC BY-NC-SA 4.0 — non-commercial only, and the classic artwork is CC BY-SA 4.0. Fine for a frame in your kitchen; read the terms before anything commercial.
  • Geography decides the payoff. Outside the well-covered regions, more of your birds will come up without an illustration.
  • It needs a decent mic. Detection quality is an audio problem before it’s a model problem.

Why it’s in this knowledge base

Because it’s the clearest small example of the pattern the rest of this site keeps circling: a model small enough to run on a $80 computer, doing one narrow job well, with no account, no API key, and no data leaving the building. Ollama makes that argument abstractly. A frame that knows a blue tit just landed outside makes it concrete.

Next

An agent with its own browser, terminal, and files, on your hardware → OpenMuse

Edit this page

Last updated 26 Sep 2026, 00:00 UTC. history