Self-hosting OpenStreetMap

I just had this weird need to self-host OSM for just internal use, not accessible from the Internet. These are notes on how to do it.

Source data

The easiest is to download an appropriate .osm.pbf¹ file from https://download.geofabrik.de/  . Unfortunately they are not directly usable for (at least those I’ve found) web renderers.

Downloading other whole planet

Use planetiler  in case you want to download some other regions or even the whole planet. It can generate mbtiles directly.

Converting pbf to mbtiles or pmtiles

The easiest way how to get mbtiles or pmtiles file is tilemaker :

docker run -it --rm -v $(pwd):/data ghcr.io/systemed/tilemaker:master /data/czech-republic-latest.osm.pbf --output /data/czech-republic-latest.mbtiles

What is the difference between mbtiles and pmtiles, you may ask? mbtiles format is based on SQLite and typically used by an application running on the same machine. pmtiles format is designed to be accessible via network, namely through HTTP range requests. This makes pmtiles super simple for hosting.

Downloading pmtiles directly

You can use bboxfinder  to find coordinates for bounding box of interest. Protomaps have a handy utility go-pmtiles  which can do operations directly on pmtiles and also to download selected bounding box from Protomaps’ server. You can verify available exports at https://maps.protomaps.com/builds/ .

pmtiles extract \
  https://build.protomaps.com/20260304.pmtiles \
  beroun.pmtiles \
  --bbox=49.763526,13.434906,50.299867,14.960632

I ran for just couple of seconds and downloaded created beroun.pmtiles file:

2026/03/04 18:00:29 extract.go:373: fetching 8 dirs, 8 chunks, 7 requests
2026/03/04 18:00:33 extract.go:413: Region tiles 9761, result tile entries 1149
2026/03/04 18:00:33 extract.go:422: fetching 1149 tiles, 68 chunks, 40 requests
fetching chunks 100% |████████████████████████████████████████████████████████████████████████████████████████████| (846/846 kB, 126 kB/s)        
2026/03/04 18:00:41 extract.go:578: Completed in 15.695928026s with 4 download threads (73.20369912351369 tiles/s).
2026/03/04 18:00:41 extract.go:583: Extract required 50 total requests.
2026/03/04 18:00:41 extract.go:584: Extract transferred 866 kB (overfetch 0.05) for an archive size of 828 kB

Serving the map

OpenStreetMap Web App

It is possible to directly run OpenStreetMap Web . There is a docker compose  which is a good starting point. It can also directly work with pbf data. Huge benefit is that all the metadata are kept and you can for example use search.

Importing data is about running the following command, be aware that it will need a lot of time².

docker compose -f docker-compose.yml run --rm web osmosis -verbose --read-pbf czech-republic-latest.osm.pbf --log-progress --write-apidb host="db" database="openstreetmap" user="openstreetmap" password="openstreetmap" validateSchemaVersion=no

It will work, but out of the box, the openstreetmap-website is going to use tile.openstreetmap.org for rendering raster (bitmap) tiles. This can definitely be changed in the configuration (see for example tile_cdn_url variable), but I have stopped pursuing this path at the end.

Raster Tile Server

Before abandoning the idea of using self-contained openstreetmap-website, I’ve explored raster tile servers. The easiest way seems to be overv/openstreetmap-tile-server . It is simple to run. But was not updated for 3y and more importantly was running quite slow for me.

Rendering and Serving mbtiles

maptiler/tilesserver-gl  turned out to be simple and quick way to get mbtiles rendered. This is what I use at the moment.

Rendering and Serving pmtiles

Serving pmtiles files is super easy, as any webserver with support for Range is sufficient. There is https://pmtiles.io  which can be used for viewing pmtiles, but rending is a bit odd (for example I cannot see street names).

Correct approach is going to be maplibre-gl-js  and Americana , which I didn’t have time yet. Maybe next time :)