Introduction
For: People evaluating Orchestrator Zero.
Orchestrator Zero runs AI agents on machines you control. You package agents, tools and their supporting parts as plugins; the platform places their jobs on nodes, records progress through Temporal, meters model calls and checks the limits you set.
The problem it solves
Teams want agents in production, and that usually means building a platform first:
- Scattered machines. GPU boxes, laptops, VMs and clouds, wired together with scripts.
- Fragile runs. A crash or a restart throws away hours of agent work and the money spent on it.
- No control. Nobody can say which agent runs where, with which tools, or at what cost.
Orchestrator Zero provides that platform through a server and nodes.
How it fits together
The server binary is orchestrator-zero, also called oz0. By default, oz0 server start runs management, the edge and embedded Temporal in one process. Management and edge are roles of that binary. You can run them separately when your deployment needs it.
The node binary is orchestrator-zero-node, also called oz0-node. It runs the agent runtime and plugins. A server and a node can share a machine; larger deployments put nodes wherever the work needs to run.
Nodes run the work
Any machine becomes a node with orchestrator-zero-node. A node connects out to the server’s edge role and needs no inbound network port. It runs the agents and tools from the plugins you install.
The edge is the only door in
The edge is the part of the server that nodes and apps talk to. It proxies Temporal, serves the runtime API and the LLM gateway, and checks every call against the node registry.
Temporal keeps runs durable
Every job is a Temporal workflow. Temporal records and replays workflow progress, but a lost harness session starts fresh from the job input and may need a person's approval. See Jobs and durability. Temporal runs inside the server by default.
Management keeps track
Management holds the registry, the certificate authority, tenants, sealed secrets, plugins and usage. When the roles run separately, edges can keep jobs running during a management outage using cached configuration and buffered writes. Operations that need uncached authorization wait; cache age, certificate expiry and buffer capacity limit how long work continues. See outage behavior.
A cluster can run multiple combined servers sharing PostgreSQL, or separate the server roles. Temporal can also run on your own external cluster or Temporal Cloud. See Architecture for the deployment choices and failure boundaries.
What you build, and what the platform does
| You write | The platform handles |
|---|---|
| Tools: components exposing actions through an MCP server or a flow | Environments, installation on every node, routing, restarts |
Skills: folders with a SKILL.md | Checking them, and loading them into runs under a start-up budget: an index at the start, a body when an agent needs it |
| Agents: instructions, a model, tools and limits, defined in Markdown or YAML | Placement, durability, delegation, contracts, limits, metering |
| Flows and routines: YAML steps, and the schedules, webhooks and events that start them | Durable runs, approvals, safe outputs, budgets per fire |
| Hooks and evals: functions that check calls, and cases a new version must pass | Running them on every call, and stopping a worse version before it rolls out |
| Event sources, connections and harnesses: what turns an outside delivery into events, how tools reach a system as the tenant, and outside agent runtimes | Checking deliveries, a token per call minted on the edge, one session per job |
| Your apps: anything on top, outside the platform | Jobs, results and cost through the runtime API, AG-UI, MCP or a chat channel |
Plugins have four kinds: a connector plugin connects to an outside system or agent runtime, a capability plugin supports a task or use case, a policy plugin provides checks for the organisation, and a plugin set installs a pinned combination of plugins. Your apps call agents through the platform. See What a plugin is for the parts and Concepts for the terms.
Everything specific to a vendor or a use case is a plugin. The core stays neutral: no models, tools or services are built in. The gateway speaks Anthropic's, OpenAI's and Google's APIs, and models served on your own machines go through it too.
Where things stand
Orchestrator Zero is pre-release. The documentation is available now; source downloads and release packages are coming soon. The quickstart currently requires repository access. Check Compatibility for platform requirements and integration maturity, and Known issues for operational limitations.