Deployment

Hive ships a Dockerfile and docker-compose.yml that run the REST API and the heartbeat daemon together in one container -- the agent OS in your own infrastructure.

Docker Compose (one command)

ANTHROPIC_API_KEY=sk-... docker compose up --build

• Control plane: http://localhost:8000/
• API docs: http://localhost:8000/docs

Provider keys are read from your shell (or a .env file) and passed through; set only the ones you use (ANTHROPIC_API_KEY, OPENAI_API_KEY, GROQ_API_KEY, HIVE_DEFAULT_MODEL). Agent state, logs, and workspaces persist in the hive_state volume across restarts.

Plain Docker

docker build -t hive-agentos .
docker run -p 8000:8000 \
  -e ANTHROPIC_API_KEY=sk-... \
  -v hive_state:/data \
  hive-agentos

How the image is built

• Multi-stage: a build stage installs the package (with the [api] extra) into a self-contained venv via uv sync --no-editable; the slim runtime stage copies only that venv.
• Runs as a non-root user; all writable state lives under /data (the working directory and the mount point), so the container filesystem stays read-only-friendly.
• The entrypoint runs the idempotent hive init then hive serve --host 0.0.0.0 --port 8000 --with-daemon, so a single process serves HTTP and drives agents.

Scaling notes

hive serve (without --with-daemon) is stateless and can run behind a load balancer with several replicas, all reading/writing one shared .hive/hive.db (SQLite WAL handles concurrent access). Run one --with-daemon instance (or a standalone hive start) to drive the heartbeat, and scale the API replicas separately. For heavier multi-writer loads, point the daemon and API at shared storage for /data.

Production checklist

• Set real provider API keys via secrets, not in the image.
• Put the API behind TLS and authentication (the server binds 0.0.0.0 in the container; restrict exposure at the proxy/ingress).
• Back up the /data volume (it holds the SQLite DB, event logs, and workspaces).
• Tune .hive/config.yaml (daemon.heartbeat, daemon.max_concurrent_agents) for your workload; enable approval and guardrails for untrusted use.