splat-agent

Installable remote GPU worker for SplatLab.

The worker is intentionally maintained separately from the SplatLab web/control-plane repository so GPU machines can install and upgrade it independently.

Docker (recommended)

The Docker image includes the complete runtime: CUDA, Nerfstudio, FFmpeg, FFprobe, COLMAP, and splat-agent. The host only needs Docker, the NVIDIA driver, and the NVIDIA Container Toolkit.

docker pull njlabsai/splat-agent:0.1.4

docker run --rm -it \
  -v splat-agent-config:/root/.config/splat-agent \
  njlabsai/splat-agent:0.1.4 init

docker run --rm --gpus all --ipc=host \
  -v splat-agent-config:/root/.config/splat-agent \
  -v splat-agent-data:/root/.local/share/splat-agent \
  njlabsai/splat-agent:0.1.4 start -n "major-tom"

The named volumes retain credentials and job data when the container is replaced. Use njlabsai/splat-agent:latest to track the newest published image.

Python package

uv tool install --locked .
splat-agent init
splat-agent start -n "RTX workstation" -q default

init accepts the connection JSON copied from the website's Settings screen:

{
  "server_url": "https://splat.example.com",
  "queue": "default",
  "credentials": {
    "access_key": "spl_...",
    "secret_key": "sk_..."
  }
}

Both compact and multi-line JSON can be pasted. If the JSON is omitted, invalid, or cannot be verified, init falls back to separate prompts for the SplatLab URL, access key, and secret key. start validates the local NVIDIA/Nerfstudio toolchain before registering and consuming one job at a time from the selected queue.

Use either -h or --help to display help for the main command or a subcommand.

Run the isolated package tests:

uv run pytest -q

License

splat-agent is released under the MIT License.