decocms/studio

What it solves

Organizations want their teams to use AI agents (like Cursor, Claude, custom agents) without scattering credentials or losing governance. deco Studio centralizes the infrastructure needed to roll out AI internally: one place to define agents, connect private tools via MCP, choose models, enforce access control (SSO, RBAC), track usage and cost, and audit every action. It replaces a pile of separate configs and ad-hoc setups with a single control plane.

How it works

Studio is a TypeScript based application with a Hono API and React 19 admin UI, backed by PostgreSQL, NATS, and object storage. It runs locally with Docker Compose or in production with Kubernetes. The core idea: everything (agents, tools, models) is exposed through one MCP endpoint. Agents package context, tools, and policy and can be published to the organization. Connections register MCP servers (GitHub, Slack, Postgres, etc.) through a web UI with OAuth, storing tokens in an encrypted vault. Models are interchangeable via OpenRouter or direct providers, and you can assign different models per agent or tool. Virtual MCPs compose and filter toolsets. Every call is traced with OpenTelemetry, giving per-user / per-agent cost and latency dashboards. The system supports multi-tenancy with organization/project isolation, and includes a sandboxed environment for code-execution agents.

Who it’s for

Organizations and teams that want a private, self-hosted AI workspace with strong governance — from a single team starting small to a large enterprise rollout. It suits platform engineers who want to standardize AI usage (approved models, tools, context) and give users chat-like access to agents, while keeping auditing and cost control.

Highlights

• One MCP endpoint for all agents, tools, and models — callable from Cursor, Claude Desktop, VS Code, or custom code
• Governed connections: OAuth setup, encrypted token vault, tool-level permissions, share capabilities without sharing credentials
• Interchangeable model layer: OpenRouter or direct providers, with per-agent model policies
• Virtual MCPs with strategies like full-context, smart selection, and sandboxed code execution
• Built-in observability: traces, costs, errors, latency per user/agent/connection
• SSO + RBAC via Better Auth, tool-scoped API keys, and multi-tenancy
• Self-host with Docker Compose or Helm on Kubernetes; deploy anywhere
• Extensible via TypeScript: define custom tools with runtime validation, audit logging, and tracing