Skip to main content
Self-hosted Temporal — a durable workflow engine — at temporal.augustin.ai. It’s the durable runtime layer for the agent platform: the piece that lets long-running agents survive crashes, retry failed steps with backoff, and schedule themselves. Workers run agent code; Temporal persists every step so a restart resumes exactly where it left off.

​
Deployment

A deliberately minimal install — three containers, ~1GB — ported from upstream’s docker-compose-postgres.yml to a Swarm stack pinned to debian-1.
ContainerImageRole
temporal_temporaltemporalio/auto-setup:1.29.1All-in-one server (frontend + matching + history + worker in one process), gRPC :7233
temporal_postgresqlpostgres:16-alpinePersistence and visibility store
temporal_uitemporalio/ui:2.34.0Web UI :8080, fronted by Traefik
Two choices keep this out of the enterprise-cluster weeds:
  • PostgreSQL, not Cassandra (DB=postgres12). The default auto-setup uses Cassandra, which is overkill here.
  • No Elasticsearch (ENABLE_ES=false). Postgres handles visibility (the searchable workflow index) at homelab scale — Elasticsearch would roughly triple the footprint for nothing.
There’s no service split (separate frontend/matching/history/worker deployments) and no replicas. That’s the over-engineered path; a single auto-setup process is the right altitude for one homelab.

​
Networking

The server and UI join the shared swarm overlay; the server and Postgres also share a private temporal-net overlay. Agent workers connect to the gRPC frontend at temporal:7233 over the swarm overlay — any stack on it can dial through (verified from the Mastra container). The UI talks to the same frontend internally and is exposed at temporal.augustin.ai via Traefik, with TEMPORAL_CORS_ORIGINS set so browser API calls are accepted. temporal-net is intentionally not --attachable, so ad-hoc CLI containers must attach to swarm instead.

​
Management

The temporal CLI ships inside the server image — no standalone admin-tools container runs (keeping it to three). Exec into the running task: 
CID=$(docker ps --filter name=temporal_temporal. -q | head -1)
docker exec "$CID" temporal operator cluster health  --address temporal:7233
docker exec "$CID" temporal operator namespace list  --address temporal:7233
The default namespace is created automatically on first boot. Postgres credentials live in temporal/.env.

​
Status

Up and serving — frontend SERVING, UI loads, default namespace registered. The next step is pointing agent workers (Mastra, the coding frameworks) at temporal:7233 and building the first durable agent workflow.