Video search for your home lab.
One compose file.

The open-source Docker version of Orcah indexes the video on your server or NAS and serves search in your browser. Free, AGPL-3.0.

you@homelab: ~
$ bash <(curl -fsSL https://raw.githubusercontent.com/IliasHad/orcah-studio/main/install.sh)
 
Orcah Studio setup — local AI video knowledge base
Choose a folder you own. A folder inside /home/you is recommended.
Install directory [/home/you/orcah-studio]:
Use NVIDIA GPU acceleration? [y/N]: y
Installed CUDA docker-compose.yml
Path to your video library [/home/you/Videos]: /mnt/nas/video
AI provider
1) Ollama runs on this machine, no API key
2) Gemini Google cloud, fast, free key at https://aistudio.google.com/apikey
3) Skip for now configure an AI provider later
Choose 1, 2, or 3 [3]: 1
Pulling qwen2.5:7b-instruct (large download, first run only)...
Wrote /home/you/orcah-studio/.env
Setup complete.
Start Orcah Studio now? [Y/n]:
Orcah Studio is starting. The first launch can take a few minutes while Docker downloads its images.
Open http://localhost:3745 when it is ready.
docker-compose.cuda.yml
6 services
  • web:3745
    edit-mind-web:latest
  • background-jobs:4000
    edit-mind-background-jobs:latest-gpuGPU
  • ml:8765
    edit-mind-ml:latest-gpuGPU
  • chroma:8000
    chromadb/chroma:1.3.5
  • postgres:5432
    postgres:16-alpine
  • redis:6379
    redis:7

Your library is mounted at /media/videos. Search opens on :3745.

Needs Docker; NVIDIA GPU optional.

What it needs.

There's no published minimum spec yet. These are the parts that decide whether it runs well.

Runtime
Docker with Compose; Docker Desktop on macOS and Windows
GPU
Optional. NVIDIA through CUDA; no Apple GPU from Docker
AI provider
Ollama on the host, Google Gemini, or none for now
Formats
.mp4, .mov, .avi, .mkv, .mxf (MXF is transcoded first)
Disk
Room for model downloads, the index and MXF working copies
Web UI
Port 3745 by default

Your library, mounted once.

Set HOST_MEDIA_PATH to the folder that holds your video. A NAS share mounted on the host works.

Compose mounts that folder at /media/videos. The web and ML containers mount it read-only. The background-jobs container mounts it read-write.

Add folders from Settings in the web app. New files are queued once they've stopped changing for 60 seconds. On network shares the watcher may miss changes made from other machines, so use Rescan on the folder page.

The index lives in named volumes (Postgres, Chroma, Redis, models) and a .data folder beside the compose file. Your footage is never copied there, so back up the index separately.

Before you leave it running.

The defaults are set up for a first run on one machine. Change these on a shared network.

  1. 1

    Change the database password

    .env.system ships with Postgres user and password set to user and password.

  2. 2

    Close the ports you don't need

    Compose publishes every service port to the host: 3745, 4000, 8000, 8765, 5432 and 6379. Only the web app needs to be reachable.

  3. 3

    Pin your images

    The compose files track :latest. Pin a tag if you want updates only when you choose to pull.

  4. 4

    Tune throughput

    One analysis and one transcription run at a time, with a 45-minute limit per video. Raise MAX_CONCURRENT_* if your hardware has room.

On NVIDIA, model inference uses CUDA but frame decoding stays on the CPU by default, because NVDEC decoding has stalled in real use. Set GPU_FRAME_DECODE=true only if it's stable on your hardware.

It fits next to what you already run.

Use it on its own, or wire it into the rest of your stack.

Immich

Connect your Immich server and Orcah imports your named people as face references.

REST API

Search scenes, list videos and collections, and fetch media with a bearer token from Settings.

MCP server

Search your library from Claude and other MCP clients.

Docs for the API and MCP server are in the repo.

Questions from self-hosters

Something else? Email us.

What hardware do I need?

Any machine that runs Docker. There's no published minimum spec yet. The ML container runs Whisper, YOLO, DeepFace and embedding models, so more RAM and an NVIDIA GPU make indexing much faster. GPU acceleration is NVIDIA only, through the CUDA compose file.

Can it use the GPU in my Mac?

Not from Docker. Containers on macOS can't use Apple Silicon's GPU, so the installer only offers GPU acceleration on Linux and Windows. On a Mac, the native Mac app uses the GPU. See the Mac app

Does it need an internet connection or an API key?

Indexing runs in your containers. For AI features you can use Ollama on the same machine (qwen2.5:7b-instruct by default, no key), Google Gemini with an API key, or skip AI for now. Docker needs internet to pull images and models the first time.

How do I update?

The compose files use the :latest image tags (:latest-gpu for CUDA). Run docker compose pull, then docker compose up -d. Your index lives in named volumes and the .data folder, so it survives updates.

Will new files be picked up automatically?

Yes. A file watcher queues new videos in folders you've added, once a file has stopped changing for 60 seconds. On network shares the watcher may miss changes made from other machines; use Rescan on the folder page to catch up.

docker compose up.
Then search.

Free and open source, AGPL-3.0.

Would rather not run a server? Get Orcah Studio for Mac · $199 once, no subscription, 30-day refund.