Skip to content

Self-hosted deployment

antcrew-platform ships as a single Docker image. For teams of 1–15 running on one machine, no external services are required — it uses SQLite and an in-process task queue out of the box.

Single-container quickstart

docker run -d \
  --name antcrew \
  -p 8000:8000 \
  -v ./data:/data \
  -e ANTHROPIC_API_KEY=sk-ant-... \
  ghcr.io/antcrew/platform:latest

On first start, the platform:

  1. Creates a SQLite database at /data/platform.db
  2. Generates a PLATFORM_API_KEY and saves it to /data/.credentials
  3. Logs the key prominently in the startup output

Retrieve it with:

docker logs antcrew 2>&1 | grep "PLATFORM_API_KEY="
# or read the file directly:
cat ./data/.credentials

Pass the key in the X-Api-Key header for all API calls.

What runs inside one container

Component Mode
API + dashboard FastAPI / uvicorn, single worker
Database SQLite at /data/platform.db
Pipeline execution In-process ThreadPoolExecutor (4 workers by default)
Scheduler asyncio background loop — eval and run schedules fire every 60 s
SSE fan-out In-process; all clients served by the same worker
Rate limiting Per-process in-memory sliding window

Everything that would otherwise require Redis or Celery degrades gracefully to an in-process equivalent. No data is lost; the only trade-off is that a worker restart interrupts in-flight runs (they are marked interrupted on the next start).

Environment variables

Required:

Variable Description
ANTHROPIC_API_KEY Default provider key used when no workspace BYOK key is set

Optional but recommended for any persistent deployment:

Variable Default Description
DATA_DIR /data Directory for SQLite database and generated credentials
DATABASE_URL sqlite+aiosqlite:////$DATA_DIR/platform.db Override to use PostgreSQL
ANTCREW_ENCRYPTION_KEY (unset) AES-GCM-256 key for encrypting run state at rest. Generate: python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"
BYOK_ENCRYPTION_KEY (unset) Encrypts BYOK provider keys at rest. Same generation command.
CORS_ORIGINS http://localhost:8000 Comma-separated list of allowed browser origins
ANTCREW_WORKERS 4 Thread-pool workers for in-process pipeline execution

Persisting data

Mount /data as a Docker volume to survive container restarts:

# Named volume (Docker manages it)
docker run -v antcrew-data:/data ...

# Bind mount (you control the path)
docker run -v ./data:/data ...

The SQLite database and the auto-generated PLATFORM_API_KEY both live in this directory.

Upgrading

docker pull ghcr.io/antcrew/platform:latest
docker stop antcrew && docker rm antcrew
docker run -d --name antcrew -p 8000:8000 -v ./data:/data ... ghcr.io/antcrew/platform:latest

SQLite migrations run automatically on startup — no manual steps required.

Scaling up

When you need multi-replica deployments, add PostgreSQL, Redis, and Celery. The same image supports all modes — they activate via environment variables:

docker run \
  --build-arg APP_ENV=prod \
  -e DATABASE_URL=postgresql+asyncpg://user:pass@db/antcrew \
  -e REDIS_URL=redis://redis:6379/0 \
  -e REDIS_PUBSUB_URL=redis://redis:6379/0 \
  -e CELERY_BROKER_URL=redis://redis:6379/1 \
  -e ANTCREW_ENCRYPTION_KEY=... \
  -e BYOK_ENCRYPTION_KEY=... \
  ...

See deployment.md for the full compose setup and Hetzner Terraform config.

Limitations of the single-container mode

  • Single-writer: SQLite locks on writes. Under high concurrency (>10 simultaneous pipeline runs), you may see contention. Switch to PostgreSQL when that happens.
  • No worker recovery: If the container restarts mid-run, the run is marked interrupted. Celery workers retry on failure; the in-process executor does not.
  • Rate limiting is per-process: If you run multiple uvicorn workers (--workers N), effective rate limit is RATE_LIMIT_RPM × N. Set REDIS_URL to share a single limiter.
  • SSE is in-process: Live dashboard works fine with one container. With multiple replicas behind a load balancer, a client on worker A may miss events from a run executing on worker B unless REDIS_PUBSUB_URL is set.