Introduction
What Agent Orchestration is, how its parts fit together, and where to start in this documentation.
Agent Orchestration is a control plane for AI coding agents. You describe a task; the platform finds a machine that has the project and a suitable agent, runs Claude Code, Codex, Gemini CLI, OpenCode or Aider there, recovers when the agent hits a limit or crashes, runs your own checks, and only then commits the result under your Git policy.
The parts
| Part | Where it runs | What it does |
|---|---|---|
| Control plane | Your server (self-hosted) or ours (managed cloud) | Accounts, organizations, projects, the task queue and scheduler, events, audit log, notifications, the web dashboard and the API. MongoDB holds the state; Redis is optional. |
| Workers | Your developer machines or build servers | A native process that connects outbound to the control plane, runs agents inside mapped project folders, verifies results and handles Git. Credentials stay on the machine. |
| Agents | On each worker | The coding CLIs you already use, driven through adapters. By default each uses its own login. |
| Clients | Anywhere | Web dashboard, agentctl CLI, VS Code extension, mobile app, REST API. |
Self-hosted and managed cloud run the same software. The difference is who operates the control plane.
Start here
What is verified
Every claim in this documentation comes from the core repository and its tests. Where something has only been tested against a stand-in — a fake provider API, a single-node cluster — the page says so. The full list lives in the core's known limitations.