Quickstart - Deco CMS

1) Run deco Studio

Pick the tier that fits. Studio hard-needs PostgreSQL and an S3-compatible object store; NATS is bundled by the chart; the sandbox (code-execution + previews) is an optional layer. The repo’s selfhost/ folder has ready-made scripts, example values, and the full dependency map.

Tier	How	Dependencies
Local — CLI	`bunx decocms`	embedded, on your machine
Local — Compose	`docker compose -f deploy/docker-compose/docker-compose.postgres.yml up`	all bundled (Postgres + NATS + MinIO)
Local — Kubernetes	one `helm install` of the umbrella `selfhost/examples/k8s-local`	all bundled (dev Postgres + MinIO + NATS + sandbox)
Production — Kubernetes	`helm install` the chart	external/managed Postgres + S3 + NATS + ClickHouse

Prefer a guided install? The self-host guide for coding agents ( SKILL.md ) interviews you — dependencies managed or in-cluster, secrets, sandbox — then writes an install directory you keep: an umbrella chart with pinned OCI dependencies plus your values, validated with helm template before anything is applied. The same directory installs on a laptop cluster and on a managed one, and no clone of this repo is involved. It’s plain markdown, so point any coding agent at it; in Claude Code you can install it as a skill:

Add it once (the repo is public), then tell Claude Code “install Studio”:

 curl -fsSL https://raw.githubusercontent.com/decocms/studio/main/selfhost/skills/install.sh | sh

Option A: one-command local (CLI)

 bunx decocms

Local instance with an embedded PostgreSQL — fully private. The package is published as decocms ; the binary it ships is deco .

Option B: Docker Compose (bundled everything)

Runs Studio + PostgreSQL + NATS + MinIO as containers, all wired — no external services needed. Multi-arch images (runs native on Apple Silicon).

 git clone https://github.com/decocms/studio.git
cd studio
docker compose -f deploy/docker-compose/docker-compose.postgres.yml up
open http://localhost:3000     # sign up — first user becomes org owner

Override the dev defaults (secrets, credentials) via a .env next to the compose file. More details: Local: Docker Compose

Option C: run from source

 git clone https://github.com/decocms/studio.git
cd studio
bun install
bun run dev

Option D: Kubernetes

The official chart ( deploy/helm/studio ) is lean: it deploys Studio + NATS and expects external PostgreSQL, object storage (S3), and (optionally) ClickHouse.

For a local cluster (Rancher Desktop / kind / minikube), an umbrella chart ( selfhost/examples/k8s-local ) composes the lean chart + sandbox layer as Helm dependencies and adds throwaway Postgres + MinIO — so one helm install brings up the whole stack, wired, including the sandbox:

 helm dependency build selfhost/examples/k8s-local
helm install deco-studio selfhost/examples/k8s-local -n deco-studio --create-namespace
kubectl -n deco-studio port-forward svc/deco-studio 8080:80   # → http://localhost:8080 (reliable on any cluster)
# studio.localhost also works IF the LoadBalancer claimed :80 — on stock Rancher Desktop k3s Traefik owns :80, so use the port-forward

Use the release name deco-studio . ./selfhost/scripts/local-k8s.sh wraps this (plus rollout wait and uninstall ).

For production, point the chart at managed services and use external secrets

autoscaling. See Kubernetes: Helm Chart.

The full picture — tiers, the batteries-included vs. swappable-dependency model, the sandbox (egress lockdown + previews), local Kubernetes runtimes, and the production posture — lives in selfhost/README.md .

2) Create a Connection

When you open a new organization, deco Studio will prompt you to:

Browse Store or
Create Connection directly - provide the MCP URL (HTTP endpoint) and any required credentials

The Store is a catalog of Connections you can connect to. By default, the Store points to the official MCP Registry.

3) Create an organization and invite members

Go to Members
Invite teammates
Assign roles (Admin/User) as needed

4) Create an Agent (virtual server)

Go to Agents
Click Create Agent
Choose Include (only selected connections/tools) or Exclude (everything except the exclusions)

5) Generate an API key for a client

Create an API key with the minimum tool permissions needed (see API Keys).

6) Verify monitoring

Use Monitoring to confirm:

tool invocations are being logged
latency + error rates show up
filters (time range, tool name, connection, agent) work as expected

The fastest smoke test is to call a tool through an agent endpoint (so you exercise aggregation + proxy + monitoring).