Skip to content

Docker

Atlas ships container images alongside the binary. Like the binary, the image is one artifact with the role chosen by subcommand — the entrypoint is atlas, so docker run … <image> up runs atlas up. Images are published to ghcr.io/orchestra-hq/atlas.

Tag Base Engine runtime Use it for
:slim, :latest, :<version> debian-slim Downloaded on first run (like the binary) llama.cpp, laptops/CPU, the default. Multi-arch.
:cuda, :<version>-cuda nvidia/cuda vLLM venv baked into the image GPU hosts serving with vLLM. linux/amd64 only.
Terminal window
docker run --rm -p 8080:8080 -v atlas-state:/var/lib/atlas \
ghcr.io/orchestra-hq/atlas:slim \
up --model qwen3-0.6b --addr 0.0.0.0:8080

Atlas prints a default API key on first start (read it from docker logs), then:

Terminal window
ANTHROPIC_BASE_URL=http://localhost:8080 ANTHROPIC_API_KEY=<key-from-logs> claude

Requires the NVIDIA Container Toolkit so the container can see the GPU (--gpus all):

Terminal window
docker run --rm --gpus all -p 8080:8080 -v atlas-state:/var/lib/atlas \
ghcr.io/orchestra-hq/atlas:cuda \
up --engine vllm --model Qwen/Qwen3-8B --addr 0.0.0.0:8080

All runtime state — downloaded engine runtimes, model weights, logs, and the keys/usage database — lives under /var/lib/atlas. Mount a volume there so it survives restarts and you only download a model once. The gateway listens on 8080; the default command binds 0.0.0.0:8080.

Keys and the usage ledger live in the state-dir database, so they survive restarts when the volume is mounted. Manage them with atlas keys / atlas usage (docker exec <container> atlas keys create …). See API keys, Usage, and TLS.