Skip to content

Run AI agents on infrastructure you control.

Your agents, your machines. Your policy decides which calls run, wait for a person or are refused.

A planet with your machines as lights on the ground, the edge as satellites they connect to, Temporal and management in orbit, and work travelling between them.

Studio

From a sentence to real work

Describe an agent, and Studio AI drafts it as a plugin. Save it, install it where it can reach, and the next incident runs through it.

  1. Start from an empty draft and say what the agent should do, in a sentence.

  2. It drafts the agent and the connectors it calls, and your policy gives each function a rule: reads are allowed, updates ask a person, deletes are refused.

  3. Save it with a message, as you would a commit. A revision never changes.

  4. Install it on the nodes labelled site=stockholm: one node, in the office with the problem.

  5. A real incident starts the job. The delete is blocked, the update waits for anna.k, and the incident is resolved.

Describe it

Start from an empty draft and say what the agent should do, in a sentence.

Studio/netopsDraft · no changesSave revision

An empty draft

Agentnew

dns-triage

Finds why one office gets a different DNS answer, and fixes it

model alias defaultpydantic-ai

Connectornew

oz0_net_diag

Probes from the node it runs on

  • dns_lookupreads
  • tls_checkreads

Connectornew

dnsyour connector

Records on the office resolvers

  • get_recordreads
  • update_recordAsk
  • delete_recordRefuse

Connectornew

servicenow

Writes up and resolves the incident

  • add_work_noteallowed by policy
  • resolve_incidentallowed by policy
agents/dns-triage.yamlthe file Studio AI wrote, in the demo’s netops plugin
name: dns-triage
runtime: pydantic-ai
model:
  use: default
tools:
  - net-diag/oz0_net_diag.dns_lookup
  - dns.update_record
  - dns.delete_record
# Its probes must see the network as the Stockholm office does.
requires:
  labels: {site: stockholm}
limits:
  max_cost_usd: 0.5

For builders: one file per agent

The agent names a model alias, its tools and where it may run. Your tool policy, not the agent, decides each call.

Agents run on Pydantic AI, the default, or the OpenAI Agents SDK, or as a harness session such as the Claude Agent SDK, OpenCode or goose. A LangGraph graph or a CrewAI crew runs as your own agent through the harness protocol. Tools are MCP servers, in any language, and oz0 job trace shows each call with what it cost.

How it runs

Where each part runs

Nodes are your machines, in as many sites as you have. They dial out to the edge, which checks every call. By default one server runs management, the edge and Temporal; you can run them apart.

The fleet
site=stockholm
site=gothenburg
site=london
Hover, tap or tab to a part.

Nodes

Your machines, in any site: servers, VMs, laptops or GPU workstations. They connect out to the edge, so they need no inbound ports.

What it holds
The agent runtime, its plugins as separate processes (sandboxed on Linux) and a certificate of its own. No model keys: model calls go through the edge.
When it is down
Finished steps stay in the job’s history. A model call in flight is tried again; the job’s next tool call waits for the node to come back.

When a node is lost

Finished steps come from the job’s history and never run twice. A model call in flight is tried again. A tool call waits for its node; a failed one is retried only if it is read-only or idempotent.
  1. planrecorded
  2. dns_lookuprecorded
  3. model callnode lost
  4. model callattempt 2
  5. update_recordwaits for its node

Rules your security team can read

  • Allow, ask or refuse

    Your tool policy decides each call by what the tool does. Destructive and unknown calls ask a person by default; a refused call never runs, and the agent is told why.

  • Nobody answering denies the call

    Every approval has a deadline, 24 hours unless you set one. Notification rules reach people by email, a webhook or a Teams, Slack or Discord channel, and escalate before it.

  • The decision covers the exact arguments

    The approver sees the exact arguments, in the console, the CLI or your app, and approves those only. Their reason goes back to the agent.

  • Your sign-in, your roles

    People sign in through your identity provider over OpenID Connect, such as Okta, Microsoft Entra ID or Keycloak, and its groups can set their roles. Operators, admins and owners decide approvals.

Approvalsnetops · waiting

netops/netops.set_isis_metric needs approval · requested by netops/noc-triage AI agent

asked 46 min ago · denied in 23 h if no one decides

level L2Escalates in 3 steps before the deadline. Next step in 5 h.

First write call in this job (earlier calls were only read)This call can do more than anything the job did before. Check that the task needs it.
{
  "link": "STO-GOT",
  "metric": 1000
}

Decision applies to arguments sha256:a11808700af2

ApproveEdit and approveDeny

A NOC agent asks to drain a backbone link in the live demo. Nobody deciding denies it.

Tool policynetops

#Applies toDecisionDeadlineFrom
17dns.delete_recordevery callRefuse–Added by handStale records are corrected with an update a person approves, never deleted
18dns.update_recordevery callAsk24 h (default)Added by handA DNS change is approved by the DNS owners
The two rules the run at the top of this page meets, each with its reason.

Watch the introduction

Why agents should run on your own machines, and the three steps to get there: connect your machines, build your agents, run and watch. With captions.

Watch it on YouTube
The film’s names and numbers are sample data. It plays from YouTube once you press play.
Read the transcript

AI agents are moving from answering questions to doing the work. They change configs, close tickets, and merge code.

But most agents run somewhere you can’t see, holding keys to your systems. And when one acts, it’s hard to say who allowed it, or what it touched.

So we built Orchestrator Zero. Your agents run on your own machines, and every tool call they make goes through rules you set.

Start a server, and join your machines with one line each. Label them by what they can reach.

Describe an agent in Studio and let Studio AI draft it. Or write it as one file in the framework you already use. Then decide what it may do. Reads? Allowed. Changes? A person signs off. Deletes? Refused.

Jobs run on the machines that can reach the problem. You watch them live. Step in when one asks. And every step’s evidence stays on the record.

We think agents will run like the rest of your infrastructure: next to your systems, on the models you choose, with every action on the record. That’s why Orchestrator Zero is open source, and why nothing in it calls home.

Orchestrator Zero. AI agents on infrastructure you control.

Built on Temporal. Connect the tools your team uses.

Every job is a Temporal workflow. Temporal runs inside the server by default, or you point Orchestrator Zero at your own cluster.

Apache 2.0 core. Your infrastructure.

The core is Apache 2.0 licensed. Its source is in the final stage of security testing before it is published. You operate the deployment and pay for your infrastructure and model services.

The northern lights over northern Europe at night, seen from the International Space Station, with city lights below.

What would you give an agent first?

Explore a task your team already does, then decide what access, evidence and review it needs.

Rev, our arctic fox, is named for revontulet, the Finnish word for the northern lights: fox fires.