decocms/studio

Open-source control plane for your AI agents. Connect tools, hire agents, track every token and dollar

deco Studio – Private AI‑agent workspace for teams

What it is – An open‑source, TypeScript‑first platform that lets an organization run its own internal AI agents. It provides a single Model Context Protocol (MCP) control plane where agents, tools, and model providers are registered, authenticated, and observed. Teams can create “agents” (bundles of context, tools, and policies), connect them to private services (GitHub, Slack, databases, etc.) and route model calls through a configurable, interchangeable model layer (OpenRouter, Anthropic, OpenAI, Google, Ollama, …). All of this can be run locally, self‑hosted, or in the cloud with SSO, RBAC, audit logs and cost accounting built‑in.

Core concepts

Concept	What you get
Agents	Publishable units that package instructions, files, and a set of tools. Each agent exposes its own MCP endpoint that can be called from any client (Cursor, Claude Desktop, VS Code, custom code).
Connections	One‑click registration of private services (GitHub, Slack, Postgres, etc.) via a web UI. Tokens are stored in an encrypted vault and access is granted per organization, role, or agent.
Models	A pluggable model router – you can point agents to OpenRouter, Anthropic, OpenAI, Google, or any compatible provider, even per‑tool.
Projects	UI containers that group agents and connections around a business goal, automatically adapting the sidebar to the relevant tools (CMS, analytics, etc.).
Virtual MCPs	Composite MCP endpoints that expose a curated subset of tools (full‑context, smart selection, on‑demand sandboxed code execution).
Observability	OpenTelemetry‑backed traces that record user, agent, model, tool, latency, token usage and cost. Dashboards let you break down spend by agent, connection, team, or organization.
Access control	SSO via Better Auth (OAuth 2.1) plus fine‑grained RBAC and API‑key scopes.
Multi‑tenancy	Isolation of config, credentials and audit logs per organization/project.
Event bus	Pub/sub with at‑least‑once delivery, scheduled/cron jobs, and sandbox‑operator integration.

Quick start (local development)

# Run the bundled CLI – it installs, starts a PostgreSQL instance and serves the UI
bunx decostudio
# or run from source
git clone https://github.com/decocms/studio.git
bun install
bun run dev   # http://localhost:4000

The stack runs on a single Bun process with an embedded PostgreSQL database, so a laptop is enough for experimentation.

Production deployment options

Target	How to deploy
Docker (single‑process)	`docker compose -f deploy/docker-compose/docker-compose.yml up` (embedded PostgreSQL)
Docker + external DB	`docker compose -f deploy/docker-compose/docker-compose.postgres.yml up`
Kubernetes/Helm	`helm install deco-studio oci://ghcr.io/decocms/chart-deco-studio --version <ver> -n deco-studio --create-namespace` – you provide PostgreSQL, optional NATS (event bus), ClickHouse + OTel collector (tracing & analytics), and the sandbox operator for isolated agent runtimes.
Bare‑metal/Bun	`bun run build:studio && bun run --cwd=apps/api start`

All deployments expose a single MCP endpoint that all agents, tools and models share.

Tech stack (as described in the repo)

Runtime: Bun (or Node) + TypeScript
API framework: Hono
Frontend: Vite + React 19 + Tailwind v4 + shadcn UI components
Database: Kysely abstraction over embedded PostgreSQL or external PostgreSQL
Auth: Better Auth (OAuth 2.1 + API‑key support)
Observability: OpenTelemetry (traces, metrics, cost aggregation)
Protocol: Model Context Protocol (MCP) – a unified way to call agents, tools and models
Packaging: Multiple npm workspaces (apps/*, packages/*) for API, docs, desktop (Tauri), UI, bindings, sandbox, etc.

Typical use cases

Secure internal AI assistants – give engineering, product or support teams a chat interface that can run code, query internal databases, or call private APIs without exposing credentials.
Enterprise model governance – enforce which models are allowed per project, switch providers without changing agent code, and track spend per team.
Tool‑driven automation – register internal services (CI pipelines, ticketing systems, data warehouses) as MCP tools and let agents orchestrate them.
Multi‑team rollout – start with a single pilot team, then expand organization‑wide while re‑using the same vault‑managed connections and audit logs.

Roadmap highlights (from the README)

Agent marketplace for discovery and composition
Declarative planning engine
Cost analytics with spend caps
Remote‑browser access improvements
Live tracing debugger
Guard‑railed workflow orchestration

License & community

License: MIT
Community: Discord, GitHub issues, and a good first issue label for newcomers. The repo includes contribution guides, testing docs and a list of maintainers.

Bottom line – deco Studio is a full‑stack, self‑hostable platform that abstracts away the plumbing required to run private AI agents at scale. It lets organizations keep data and credentials on‑premise while still taking advantage of modern LLM providers, with built‑in observability, access control and cost tracking.