Run the server

On the host system:

• Docker v29+ [docs]
• (Optional, recommended) Just v1.46+ [docs]
• (Only for Intel® XPU) the latest driver suitable with your HW [docs]
• (Only for NVIDIA GPU) NVIDIA driver and the NVIDIA Container Toolkit [docs]

Geti Docker images can be built from source using the Dockerfile in the application directory. This can be useful if you want to customize the application or optimize it for a specific use case that is not covered by the pre-built images.

The instructions below use just to simplify the build process, but you can also build the image manually with docker build if you prefer.

From the application directory:

# Choice of accelerator: 'xpu', 'cuda', or 'cpu'
just build-image --accelerator xpu

Run just --usage build-image to see all available build options.

By default, the container generates a self-signed certificate at startup and serves over HTTPS on the configured port. For production deployments, mount your certificate and private key into the container using --volumes, then point --certfile and --keyfile to the in-container paths:

just run-image --accelerator xpu \
    --volumes "/path/to/certs:/certs:ro" \
    --certfile /certs/server.pem \
    --keyfile  /certs/server-key.pem

The cert directory is mounted read-only and is separate from the data volume - it is never modified by the container.

Geti uses WebRTC for the real-time inference streaming visualization in the UI. WebRTC requires the browser to establish a direct connection to the backend's media server: when the application is deployed in a restricted network environment, such as behind a corporate firewall or NAT, the connection may fail due to the inability to traverse the network boundaries. To address this issue, you can set up a TURN server that acts as a relay between the browser and the backend, allowing the WebRTC traffic to pass through the restricted network.

First, set up a TURN server by running just run-coturn from the application directory. This command will launch a Coturn TURN server in a Docker container with default credentials.

Next, pass the --coturn option when launching the Geti container via just run-image. That's it, now you should be able to view the predictions in the 'Inference' page of the UI.

Later, when you no longer need the TURN server, stop it with just stop-coturn.

The Geti application uses a Docker volume named geti-data to persistently store all datasets, models, and other objects. You can browse the contents of this volume by running a temporary container that mounts the volume and lists the files.

# List the contents of the root directory in the `geti-data` volume
docker run --rm -v geti-data:/data alpine ls -l /data

# List the model files of a specific project (replace <PROJECT_ID> with the actual ID)
docker run --rm -v geti-data:/data alpine ls -l /data/projects/<PROJECT_ID>/models

# List the media files of a specific project (replace <PROJECT_ID> with the actual ID)
docker run --rm -v geti-data:/data alpine ls -l /data/projects/<PROJECT_ID>/dataset

When running Geti with Docker, all logs are stored in the geti-logs Docker volume. You can view these logs by running a temporary container that mounts the volume and prints the log files to the console.

These examples use jq to format the JSON logs; install jq on the host or omit the | jq -r '.text' part to see the raw JSON output. Application logs:

# Print the logs of the application container to the console
docker run --rm -v geti-logs:/logs alpine cat /logs/app.log | jq -r '.text'

# Or save the logs to a file for easier browsing
docker run --rm -v geti-logs:/logs alpine cat /logs/app.log | jq -r '.text' > geti-logs.txt

Job logs:

# List the available job logs
docker run --rm -v geti-logs:/logs alpine ls -l /logs/jobs

# Print the logs of a specific job to the console
docker run --rm -v geti-logs:/logs alpine cat /logs/jobs/<job_type>-<job_id>.log | jq -r '.text'

Logs of other worker processes:

# Print the logs of the inference pipeline stream loader
docker run --rm -v geti-logs:/logs alpine cat /logs/workers/streamloader.log | jq -r '.text'

# Print the logs of the inference worker
docker run --rm -v geti-logs:/logs alpine cat /logs/workers/inference.log | jq -r '.text'

• Just v1.46+ [docs]
• (Only for Intel® XPU) the latest driver suitable with your HW [docs]
• (Only for NVIDIA GPU) NVIDIA driver and the NVIDIA Container Toolkit [docs]
• Node.js v24.2+ [docs]

To run the server, use the run-server command after initializing the environment with venv:

# From the repo root
cd application/backend

# Initialize the environment with the appropriate accelerator support (cpu, xpu, or cuda)
just venv --accelerator xpu

# Run the server
just run-server

Run just --usage run-server for a full list of options for running the server. Notably, by passing the option --setup-demo, the application will be pre-populated with demo data, including sample datasets and pre-trained models.

After running the server, build and launch the UI in a separate terminal:

# From the repo root
cd application/ui

# Install dependencies and build
npm install
npm run build

# Start the UI
npm run start

After the UI starts, you can access the Geti web application at http://localhost:3000 (assuming default settings).

The OpenAPI specification for the Geti REST API can be generated with the generate-api-spec command:

# From the repo root
cd application/backend

# Generate the OpenAPI spec and save it to a custom location
just gen-api-spec --output-path="openapi.json"