Introduction
What Org in a Box is, who it's for, and how it works.
Org in a Box: Enterprise AI Workforce Platform
Supported today: the checked-in, demo-ready path is local Docker Compose with one sandbox container, explicit
runtime apply, and the web or TUI client. The Azure docs in this repo describe a preview deployment path, not a fully wired remote sandbox pool.
Org in a Box (OIAB) gives every employee in your organisation their own team of autonomous AI agents. Agents live inside sandboxed runtimes, communicate over channels your team already uses (Slack, Teams, email), and execute long-running tasks while you focus on higher-level work.
Key Concepts
| Concept | Description |
|---|---|
| Sandbox runtime | In the local demo path, a shared Ubuntu sandbox runs opencode serve and reloads when you apply a new runtime bundle. |
| Channel Gateway | Inbound messages from Slack / Teams / email are routed through a gateway into the job queue, which dispatches an agent-turn job. |
| Job Queue | PostgreSQL-backed queue with SELECT FOR UPDATE SKIP LOCKED. LISTEN/NOTIFY wakes the drainer immediately when a new job lands. |
| Memory | pgvector stores 1536-dim embeddings per memory. system.transform injects the top-K semantically-similar memories into each agent turn. |
| Orchestrator | An LLM decomposes complex tasks into parallel sub-tasks with a dependency graph. Each sub-task runs in its own session. A synthesis step combines all results. |
| Learning Loop | After each session, a reflection LLM extracts a trajectory hash. Three similar trajectories auto-propose a reusable skill. Admins promote skills to make them available globally. |
| RBAC | Four system roles (owner / admin / member / viewer) exist in the schema today. Browser SSO currently creates tenant users and assigns the default role in-app; automatic group-to-role mapping is not fully wired in the checked-in callback flow yet. |
When to Use OIAB
- You want every employee to have an AI agent accessible through existing channels.
- You need audit trails of tool calls, memory reads, and auth events.
- You want a local or self-hosted compose stack you can demo and iterate on quickly.
- You want parallel multi-task execution, not just a single chat thread.
Navigation
- Quick Start: Up and running in minutes with Docker Compose.
- Features: Deep-dives on every major feature.
- Architecture: Database schema, component topology, data flows.
- API Reference: Every REST endpoint with request/response shapes.
- Deployment: Local Compose guidance plus an honest Azure preview status.
- Changelog: Release notes for demo-verified changes.