AmberSight
Docker app from UnraidUser-7366147911's Repository
AmberSight is a self-hosted camera recording server: it records RTSP cameras, detects motion, checks it with built-in AI object detection, and sends alerts.
Overview
The data folder holds the SQLite database. It is best kept off Unraid's /mnt/user layer (FUSE): on Unraid 6.12+ set the appdata share to Exclusive access, or point Data at /mnt/your pool/appdata/ambersight/data. To record to an array share, map it to /recordings (it must be writable by PUID:PGID, 99:100 by default; an existing share is fine) and choose /recordings in Settings Recording Drives & folders.
Bridge networking works for cameras added by address. Camera discovery (ONVIF) and battery Reolink cameras need host networking. Hardware video decode on Intel or AMD: add --device=/dev/dri to Extra Parameters (Advanced View). For an NVIDIA card use the :cuda tag with the Nvidia Driver plugin and --runtime=nvidia (or --gpus all) in Extra Parameters.
Readme
View on GitHubRunning AmberSight in Docker
A prebuilt image is published with each release:
| Image | For |
|---|---|
ghcr.io/huskerminion/ambersight:latest |
the newest release, amd64 and arm64 (such as a Raspberry Pi 5) |
ghcr.io/huskerminion/ambersight:<version> |
one release, e.g. 0.11.0 |
ghcr.io/huskerminion/ambersight:cuda |
the newest release with the AI on an NVIDIA card, amd64 only (also <version>-cuda) |
It holds the AmberSight program, ffmpeg and the Intel/AMD video drivers; your settings and recordings live in two folders you mount. You can also build the image yourself from this repository (Building).
Unraid: see unraid/README.md.
Quick start
Save docker/compose.yaml in a folder of its own and
run, in that folder:
docker compose -f compose.yaml up -d
(From a clone of this repository: docker compose -f docker/compose.yaml up -d.)
Then open http://<this-computer>:8090.
Add the first user (Settings > Users & security) before anything else. Until then, as with any new install, anyone who can reach port 8090 can watch the cameras and change every setting. With host networking that is every network the computer is on. Don't forward the port on your router; use a VPN such as Tailscale or WireGuard to reach it from outside.
Edit compose.yaml first if you want the settings and recordings somewhere
other than config and data next to it, and set TZ to your time zone.
Without Compose:
docker run -d --name ambersight --restart unless-stopped --stop-timeout 30 --network host \
-e TZ=America/New_York \
-v /srv/ambersight/config:/config -v /srv/ambersight/data:/data \
--device /dev/dri ghcr.io/huskerminion/ambersight:latest
Folders
| In the container | What is in it |
|---|---|
/config |
config.toml, written with starter settings on the first start |
/data |
the database, recordings, AI models, backups |
Keep /data on a local disk: the database is SQLite and should not be on a
network share. To record to a bigger or separate drive, mount it as another
folder (for example -v /mnt/cameras:/recordings) and choose it in
Settings > Recording.
AmberSight runs as user and group 1000. Set PUID and PGID to run it as
another user. When /config or /data belongs to someone else (the first
start, a changed PUID), the container gives that folder and everything in
it to that user when it starts. Files copied in later as root, into a folder
that already belongs to the user, are not changed: fix their owner by hand.
/recordings is handed to that user only when it is new: owned by root (what
Docker makes of a host folder that did not exist) and empty. Any other folder
you record to, including an existing share full of recordings, is left as it
is and must already be writable by PUID:PGID. Other new, empty mount points
can be named in RECORDINGS_DIRS (separated by colons, for example
/cams1:/cams2) to get the same treatment; a symlink is left alone. When
AmberSight cannot write to a recording folder, the health page says so, and a
camera recorded through ffmpeg shows it as its last error too.
Use PUID/PGID rather than --user: with --user the folders and the
video devices' groups are not set up. PUID=0 runs AmberSight as root, for
rootless Docker or Podman, where root in the container is your own user.
On Unraid, use PUID=99 and PGID=100 (nobody:users, who own the shares).
TZ sets the time zone (for example America/Chicago); schedules and quiet
hours use it. Without it the container runs on UTC.
Network
Host networking (network_mode: host, or --network host) is the
simplest and is what the examples use. It is needed for:
- Find cameras: ONVIF discovery uses multicast, which does not reach a container on Docker's own network.
- Battery Reolink cameras (Argus): they are found with a UDP broadcast, and their motion push arrives on port 80 (see the install guide).
On Docker's own network (bridge), publish the port (-p 8090:8090) and add
cameras by address. RTSP, ONVIF events, PTZ and everything else work the same.
Behind a reverse proxy or tunnel (nginx, Caddy, Traefik, cloudflared,
Tailscale Funnel), every request reaches AmberSight from the proxy, so one
visitor's failed sign-ins would lock out everyone. Tell AmberSight the proxy's
own address in config.toml and it takes each visitor's address from
X-Forwarded-For:
[server]
trusted_proxies = ["172.17.0.1"] # the proxy as the container sees it (Docker's bridge gateway)
List only the proxy, never a LAN or internet range: anyone in a listed range
could claim any address. The proxy must add the visitor to X-Forwarded-For
(Caddy, Traefik, cloudflared and Tailscale do; nginx needs
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;).
Ports, if you publish them one by one:
| Port | For |
|---|---|
| 8090/tcp | the web interface, API and live video (HTTP or HTTPS) |
| 1935/tcp | only if cameras push RTMP to AmberSight |
| 80/tcp | only for battery Reolink cameras' motion push |
Port 80: AmberSight does not run as root (unless PUID=0), so listening on port 80 needs
--sysctl net.ipv4.ip_unprivileged_port_start=80 on a bridge network. With
host networking that setting is the host's own.
Hardware video decode
Pass /dev/dri through (--device /dev/dri) for Intel or AMD hardware
decode (VAAPI); the container joins the groups that own those devices by
itself. Motion detection then decodes on the graphics card by itself: at
start AmberSight test-decodes a short H.264 and H.265 sample on each render
node and uses the first that decodes it right (Status > Cameras shows GPU or
CPU per camera; a card that fails is named in Status). Without /dev/dri
everything decodes on the processor, as before.
To choose per camera (camera settings > Stream > More settings > Hardware
decoding, or in config.toml):
hwaccel = "off" # the processor
hwaccel = "vaapi" # VAAPI on the first render node that passed the test
hwaccel = "vaapi:/dev/dri/renderD129" # a particular one, when there are several
hwaccel = "qsv:/dev/dri/renderD128" # Intel Quick Sync
ls -l /dev/dri/by-path on the host shows which renderD belongs to which
graphics card (an NVIDIA card's node does not do VAAPI).
AI on an NVIDIA card
The plain image runs the built-in AI on the processor. For an NVIDIA card,
use the CUDA image instead (:cuda, built from Dockerfile.cuda: NVIDIA's
CUDA 13 + cuDNN 9 runtime libraries with AmberSight and its NVIDIA add-on; it
is large, about 4 GB, and amd64 only):
docker run -d --name ambersight --restart unless-stopped --stop-timeout 30 --gpus all --network host \
-v ./config:/config -v ./data:/data ghcr.io/huskerminion/ambersight:cuda
The host needs NVIDIA's driver and the
NVIDIA Container Toolkit
(--gpus all passes the cards in). With compose, use
image: ghcr.io/huskerminion/ambersight:cuda and add:
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
Settings > Detection & alerts > Objects (AI) then shows Running on: NVIDIA
.... The AI's memory pool on the card is capped at 1.5 GB (gpu_memory_mb
in [ai]). The plain image given --gpus all finds the card but has no CUDA
libraries: it stays on the processor, downloads nothing, and says why in
Settings (libcudart.so.13 not found).
Recordings whose files are gone
After moving recordings or replacing a drive, Settings > Recording > Drives & folders can find and remove the ones whose files are gone. From the command line, run it as the container's user (not root):
docker exec -u 1000:1000 ambersight ambersight storage prune-missing --config /config/config.toml
docker exec -u 1000:1000 ambersight ambersight storage prune-missing --config /config/config.toml --apply
(use your PUID:PGID if you set them; the first line only reports). See the install guide's
"Moving recordings or moving to a new computer".
Updates
The built-in updater does not install updates in a container (an update would be lost when the container is recreated); it still tells you when a new version is out. To update, pull the new image and recreate the container:
docker compose pull && docker compose up -d
(add -f compose.yaml, or -f docker/compose.yaml in a clone, when you are
not in the folder that holds it). Without Compose: docker pull ghcr.io/huskerminion/ambersight:latest, then remove the container and run it
again with the same docker run line. An image you built yourself: git pull
and docker compose -f docker/compose.yaml up -d --build.
:latest and :cuda always point at the newest release; use a version tag
(e.g. ghcr.io/huskerminion/ambersight:0.11.0) to stay on one release and
update when you choose. Settings and recordings stay in the mounted folders.
Restarts and stopping
Turning HTTPS on or off asks for a restart (Restart now): AmberSight
exits and Docker starts it again, so keep restart: unless-stopped (or
--restart unless-stopped). Without a restart policy it stays stopped until
you start it.
On stop, AmberSight finishes the recordings and alerts in progress, which can
take longer than Docker's default 10 seconds. The examples allow 30
(stop_grace_period: 30s, --stop-timeout 30).
Building
To build the image from this repository instead of pulling it:
git clone https://github.com/HuskerMinion/AmberSight.git
cd AmberSight
docker build -t ambersight . # or: -f Dockerfile.cuda -t ambersight:cuda
With Compose, replace the image: line in docker/compose.yaml with the
commented build: .. and image: ambersight lines, and run
docker compose -f docker/compose.yaml up -d --build. The first build takes
several minutes.
The image builds with plain docker build; BuildKit is not needed. The
build stage compiles AmberSight from source (a couple of minutes on a fast
machine, longer on a Raspberry Pi), and the final image is Debian 13 slim
with ffmpeg. It builds for the machine it is built on.
Categories
Related apps
Explore more like this
Explore allLinks
Details
ghcr.io/huskerminion/ambersight:latestRuntime arguments
- Web UI
http://[IP]:[PORT:8090]/- Network
bridge- Shell
sh- Privileged
- false
- Extra Params
--stop-timeout 30
Template configuration
Web interface, API and live video (HTTP, or HTTPS once turned on in Settings).
- Target
- 8090
- Default
- 8090
- Value
- 8090
Optional. Only used when cameras push RTMP to AmberSight; you can remove it otherwise.
- Target
- 1935
- Default
- 1935
- Value
- 1935
config.toml, written with starter settings on the first start.
- Target
- /config
- Default
- /mnt/user/appdata/ambersight/config
- Value
- /mnt/user/appdata/ambersight/config
Database, recordings (unless you add a recordings folder), AI models and backups. The database is SQLite and runs best off the /mnt/user (FUSE) layer: on Unraid 6.12+ set the appdata share to Exclusive access, or use /mnt/your pool/appdata/ambersight/data.
- Target
- /data
- Default
- /mnt/user/appdata/ambersight/data
- Value
- /mnt/user/appdata/ambersight/data
Optional. A share for recordings, e.g. /mnt/user/cameras. After the first start, choose /recordings in Settings > Recording > Drives & folders. The share must be writable by PUID:PGID (99:100 by default); an existing share is fine. Leave empty to record into the data folder.
- Target
- /recordings
User AmberSight runs as; it owns the config and data folders (99 = nobody on Unraid).
- Default
- 99
- Value
- 99
Group AmberSight runs as (100 = users on Unraid).
- Default
- 100
- Value
- 100