Fitberg

Self-hosted training log for FIT files. Imports your files, keeps them unmodified, and derives training load, fitness and form, personal records, power curves, zones and a 3D map of every route.

Runs in Docker on a home server or a Raspberry Pi. No accounts, no cloud services, no network calls except map tiles — and COROS sync if you turn it on, which talks to COROS's own official service on your behalf.

Features

• Import — loose .fit files, .fit.gz, .tcx (Nike Run Club and others), a folder from a device, or an unopened export ZIP from another platform. Archives are searched recursively; files are identified by content, not extension. Re-importing the same data is a no-op.
• COROS sync — connect your COROS account once and Fitberg pulls new activities through COROS's official MCP service: real FIT files, straight from their servers, into the same pipeline as everything else. Runs about once a day while the server is up, or on demand from the Import page. COROS caps downloads at 50 activity files per day, so a first backfill of a long history takes a few days — it keeps going by itself. Off unless you connect an account; disconnecting revokes Fitberg’s access at COROS.
• Training load from power, heart rate or grade-adjusted pace, whichever the file supports, on one scale where an hour at threshold is 100.
• Fitness and form — Banister impulse-response: 42-day fitness, 7-day fatigue, form as the difference.
• Records per sport, scanned across every recorded second, so a fast 5 km inside a long run counts. Race distances for running, time-trial distances for cycling, pool events for swimming.
• Mean-maximal power curves, VO₂max estimates, HR/power/pace zones, aerobic decoupling, efficiency factor, monotony and strain.
• 3D flyover over real elevation, with a HUD and linked charts.
• Non-destructive trimming — cut a warm-up or a forgotten stop; totals and averages are recomputed, the file is untouched.
• Export — one ZIP of your FIT files, readable by any other tool.
• Optional local AI via Ollama for a weekly review and questions about your own data.

Quick start

Requires Docker.

git clone https://github.com/yonie/fitberg.git
cd fitberg
cp .env.example .env
mkdir -p data
docker compose up -d --build

Open http://localhost:8710.

Set SESSION_SECRET in .env to keep sessions across restarts, or FITBERG_OPEN_ACCESS=1 to disable the login on a trusted network.

To try it with generated data:

docker compose exec fitberg npm run seed -- 24

Data and backups

Everything is in data/, next to docker-compose.yml:

data/originals/   your FIT files, unmodified
data/fitberg.db   derived data, rebuildable

Back up data/ and you have a complete copy. The database is disposable — delete it and rebuild with reindex.

To move machines or migrate away, export the ZIP and import it elsewhere.

Configuration

All values have defaults. See .env.example.

Variable	Purpose
DATA_DIR	Where files and the database live. Compose pins this to /data.
SESSION_SECRET	Session signing key. Unset means sessions end on restart.
FITBERG_OPEN_ACCESS	1 disables authentication.
OLLAMA_URL, OLLAMA_MODEL	AI host and model. Also settable in Settings.
TERRAIN_TILE_URL	DEM tiles for terrain and hillshading.
MAP_STYLE_URL	Optional vector basemap, replacing the built-in raster style.

Screenshots

Development

Node 24 or later. No native modules.

npm run install:all
npm run build
npm start
npm test

CLI, prefixed with docker compose exec fitberg when containerised:

npm run fitberg -- status      # what is stored, and current metrics
npm run fitberg -- import ...  # files, a folder, or an export ZIP
npm run fitberg -- verify      # check stored files against their hashes
npm run fitberg -- reindex     # rebuild the database from the FIT files
npm run fitberg -- recompute   # recalculate load and fitness

Imported files are content-addressed and read-only; everything else is derived, so schema changes are applied by adding a column and reindexing. User edits — trims, notes, RPE — are keyed by an activity's sport, start time and distance so they survive a rebuild. Storage is SQLite via node:sqlite, with sample streams as typed-array blobs.

Known issues

• In 3D the route is drawn as circles, not a line: MapLibre drapes line layers onto the terrain mesh, where the route is hidden. The flat view uses a line.
• MapLibre is pinned to ^5; 6.0.0 does not render GeoJSON line layers.
• Multi-user is implemented but only tested with one user.

Licence

MIT