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.
Two variants
Section titled “Two variants”| 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. |
Quick start (slim / CPU)
Section titled “Quick start (slim / CPU)”docker run --rm -p 8080:8080 -v atlas-state:/var/lib/atlas \ ghcr.io/orchestra-hq/atlas:slim \ up --model qwen3-0.6b --addr 0.0.0.0:8080Atlas prints a default API key on first start (read it from docker logs), then:
ANTHROPIC_BASE_URL=http://localhost:8080 ANTHROPIC_API_KEY=<key-from-logs> claudeQuick start (CUDA / GPU)
Section titled “Quick start (CUDA / GPU)”Requires the NVIDIA Container Toolkit
so the container can see the GPU (--gpus all):
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 0.0.0.0:8080State and ports
Section titled “State and ports”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, usage, TLS
Section titled “Keys, usage, TLS”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.