Run every agent as a durable workflow on the machines you already own. Routed, metered and kept inside the lines.
Orchestrator Zero runs AI agents on the machines you own, in every site and every network, and keeps them in line. Package an agent once and any machine in the fleet can run it. Every step is saved, every model call is metered and nothing risky happens without a yes.
Join your machines
Servers, laptops, GPU boxes and devices in every site, with one command each.
oz0-node join --token …Install agents from Git
Agents, tools and skills come as plugins, rolled out to the nodes that need them.
oz0 plugin install https://github.com/acme/ops-agentsStart jobs
From the CLI, your apps or a schedule. Follow them live and see what each one cost.
oz0 run outage-triage "Link down in got-2"Laptops, GPU workstations and cloud VMs join with one command. They only dial out, so there is nothing to open in a firewall.
Nodes talk to the edge and nothing else. It proxies Temporal, serves the runtime API and meters every model call.
node-7f3acalls the model gateway
Each agent run is a Temporal workflow: embedded, on your own cluster or on Temporal Cloud. It survives crashes and can wait for days.
Web UI, registry, certificates, rollouts and cost reports. It runs on its own, so it can go down without stopping a single job.
Cost by agent, last 7 days
Example data
People use the web UI and the CLI, your apps call the runtime API, and other services post a webhook. Every way starts the same durable job, with the same budget, approvals and history.
Approvals 2 waiting
change-runner wants junos.commit
site got-2 · waiting 4m
ApproveDeny
containment wants host.isolate
fin-ws-042 · waiting 12m
Run pr-reviewer
{"repo": "acme/api", "pr": 42}
$2.00
Run
Start one server. Then install the node on every machine you want to use and join it with a token.
Management, edge and Temporal start together in one process.
$ oz0 server startGive it labels and an expiry. Machines that join with it are let in straight away.
$ oz0 token create --label gpu=a100 --ttl 24hRun the install script on any machine you want in the fleet.
$ curl -sfL https://oz-edge:7443/install.sh | shThe node checks the server, gets its certificate and starts taking jobs. It only ever dials out.
$ oz0-node join --server oz-edge:7443 --token K10b4e…::jt_7Qx…The machine is part of the fleet and picks up jobs that match its agents and labels.
oz0 node accept node-7f3aoz0 node block node-7f3aYou build the agents as plugins. The platform makes sure they behave, wherever they run.
Every run is a Temporal workflow. If a node dies, another node with the same agent picks it up where it stopped.
Example run: review-pr
A worse version of an agent never reaches production.
Token and cost on every model call. A job tree shares one budget, with cost per branch.
A job lands on a node with the right agent and hardware.
Tools are MCP servers and hooks speak gRPC, so any language with an SDK for either works. Install from any git repo and ref.
Claude, OpenAI and local models go through one gateway. Claude Agent SDK, Codex and OpenCode plug in as harnesses.
Run it embedded in the server, point at your own cluster or use Temporal Cloud. The same plugins and nodes work on all three.
Secrets are stored centrally and only decrypted at the edge and on the node that needs them.
Everything specific to your company ships as a plugin: a git repo with one manifest. Write it in any language that speaks MCP or gRPC.
# oz0-plugin.yaml
apiVersion: oz0/v1
name: risk-tools
version: 1.4.0
runtime: python # python | node | go | binary
requires:
secrets: [RISK_API_KEY]
tools:
- id: risk
mcp: ["python", "-m", "risk_tools.mcp"]
requires_approval: [risk.override]
- id: fetch
package: pypi:mcp-server-fetch # an existing package, no code
agents: [agents/credit-reviewer.yaml]
skills: [skills/]
hooks:
- on: pre_tool_call
match: { tools: ["risk.override"] }
on_timeout: escalate
evals: [evals/]
$ oz0 plugin install https://github.com/your-org/risk-tools --ref v1.4.0 --canary 10%Example: a team adds its own risk model and a review hook. Edge locks the commit, runs the evals, and rolls it out to 10% of nodes first.
Agent runs are long, expensive and break halfway. Temporal is the open-source engine made for exactly that kind of work, so we built on it instead of reinventing it.
It is MIT-licensed and written in Go, so it ships inside the server. Rather not run it yourself? Point Orchestrator Zero at your own cluster or at Temporal Cloud.
Durable runs on machines you own, run as a fleet from a control plane you host, with no inbound ports, budgets that stop a run and tenants built in.
| Product | Any machine you own | Durable runs | Fleet operations | Self-hosted control plane | No inbound ports | Budgets that stop a run | Tenants built in |
|---|---|---|---|---|---|---|---|
| Orchestrator Zero | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| Windmill | Yes | Yes | Tags; groups in Enterprise | Yes | Yes | Metered, no cost stop | 3 workspaces free |
| Temporal Agent HarnessExperimental | Yes | Yes | Worker list, no machine join | Yes | Yes | Tokens counted, no budget | Namespaces, access is yours |
| Hatchet | Yes | Yes | Labels, no version rollouts | Yes | Yes | Not documented | Yes |
| AgentField | Yes | Re-runs, resume is manual | Yes | Yes | Calls into agents | Caps for some harnesses | Team boundaries |
| OpenClawPersonal and team assistant | Yes | Best-effort resume | Yes | Yes | Yes | Estimates, no cost cap | One boundary per gateway |
| LangSmith DeploymentFormerly LangGraph Platform | Enterprise plan | Yes | Deployments, no routing | Enterprise plan | Server takes requests | Caps in cloud beta | Yes |
| n8n | Yes | Manual retry | No label routing | Yes | Yes | Not documented | Projects, paid plans |
| Kagent1.0 is in alpha | Kubernetes only | Lost if its node dies | Kubernetes pools | Yes | Traffic routed in | Tokens traced, no budget | Auth in Enterprise |
| Managed cloudsAgentCore, Foundry, Gemini Enterprise | Their cloud only | No step-level resume | No | No | Requests pushed in | Token quotas only | Per-user sessions |
From each product's public documentation, 2 October 2026. Partial means part of the capability, or only in a paid tier. Not documented means we found nothing either way.
Meet Rev
In Nordic folklore the northern lights are sparks from an arctic fox’s tail as it runs across the snow. In Finland they are still called revontulet: fox fires.
Rev is our fox. A small arctic fox with an aurora for a tail, it runs from machine to machine and lights up the sky with every job it finishes.
Self-hosted on your own machines. Companies run it under an enterprise agreement.
Free
For personal projects, homelabs and learning.
By agreement
For companies running agents in production. A commercial licence and direct support from the team that builds it.
Start with one server and the machines you already have.