self-hosting
Self-hostingbeta
Self-hosting
Excalibase is an open-source platform as a service. Everything that runs Excalibase Cloud runs on
your own machines too: Studio and the control plane, end-user auth, the GraphQL and REST engine,
change capture, per-project databases, storage and container hosting. The code is Apache-2.0 and
the images are public on Docker Hub (excalibase/*).
Choose an install mode
| Single host: Docker | Single host: rootless Podman | Kubernetes | |
|---|---|---|---|
| Best for | a VPS or server you control | the same, with nothing running as root | production with several machines, HA |
| Install | init.sh + docker compose | init.sh --engine podman + podman compose | helm install |
| Guide | Single host | Single host | Kubernetes |
Just looking? Try it locally in a few minutes, on localhost.
What gets installed
| Component | Does |
|---|---|
| Studio | the web console: projects, tables, SQL, permissions, users, apps, backups |
Control plane (provisioning) | creates and manages projects, the vault holding their credentials, backups, apps |
Auth (excalibase-auth) | sign-up, sign-in and tokens for the end users of each project |
API engine (excalibase-graphql) | GraphQL and REST for every project, at /{projectId}/graphql and /{projectId}/api/v1 |
Watcher (excalibase-watcher-go) | change data capture for realtime, one per project (Kubernetes) |
Function runtime (deno-runtime) | TypeScript edge functions, one runtime per project (Kubernetes) |
| Platform database, NATS | the platform's own state and its change events |
| Object store | project backups and customer files: bundled (RustFS) or your own S3-compatible bucket |
| Project databases | one Postgres per project (14, 15, 16 or 17), optionally with DocumentDB (15 to 17); CloudNativePG on Kubernetes, a container on a single host |
| Edge | HTTPS in front of Studio, the API, files and apps: Caddy on a single host, HAProxy on Kubernetes |
Requirements
Single host (Docker or Podman):
- Linux, 4 CPU and 8 GiB RAM for the platform plus a few projects.
- Docker Engine 24+ with the Compose plugin, or Podman 4.9+ with Docker Compose v2.
- A domain with
studio.,api.andfiles.records (a wildcard works) pointing at the host, and ports 80 and 443 open for Let's Encrypt. For a trial,localhostneeds neither. - For container apps: gVisor on Docker (recommended), and a wildcard record
*.apps.<domain>.
Kubernetes:
- Kubernetes 1.30 or newer. What we test: k3d for trying it, RKE2 for production.
- A CNI that enforces NetworkPolicy. Cilium 1.19+ is recommended and required for container apps (their fences are CiliumNetworkPolicies) and for edge functions' outbound allowlist by host name. Flannel alone enforces nothing: every isolation rule would be silently ignored.
- An Ingress controller: HAProxy Kubernetes Ingress (what we run) or ingress-nginx.
- cert-manager, CloudNativePG 1.30 and the Barman Cloud plugin.
- A default StorageClass with ReadWriteOnce volumes.
- Size: one node with 4 CPU / 8 GiB runs the platform and a few Free projects. The Standard tier runs 3 database instances, one per node, so it needs 3 schedulable nodes (each Standard project asks for 6 CPU and 12 GiB across them).
Both modes need, for real use, an email relay (SMTP, Amazon SES or Resend) so people other than the first admin can confirm their address. See Configuration.
Single host or Kubernetes: what differs
| Single host | Kubernetes | |
|---|---|---|
| Postgres projects, GraphQL, REST, end-user auth, permissions, SDK keys | yes | yes |
| Storage (customer files) | yes | yes |
| Container apps | yes, under gVisor on Docker; on rootless Podman without a sandbox (trusted code only) | yes, under gVisor, fenced by Cilium |
| DocumentDB projects | yes, with init.sh --documentdb | yes |
| High availability (3 or 5 instances) | no: one machine is one failure domain, so tiers with more than one instance are refused | yes, one instance per node |
| Backups | yes, to the bundled store or your bucket | yes, continuous WAL archiving to the store |
| Point-in-time recovery | not working yet: WAL archiving is not applied to single-host projects, so a restore to a point in time does nothing; restoring a backup works | yes |
| Edge functions, realtime | no | yes |
| Isolation between projects | one internal network per project | namespaces and NetworkPolicies |
| Disk limits | recorded, not enforced by the filesystem | enforced with a sized StorageClass |
Need HA on few machines? A small Kubernetes cluster (k3s or RKE2 on three nodes) is the way.