Announcement · Turf 1.0

Turf 1.0 is available now.

Turf™ is a self-hosted, autonomous infrastructure-as-code tool, built on OpenTofu™. It deploys the stack an AI factory runs on and keeps it that way, with an agent overseeing.

Turf is free for noncommercial use under PolyForm Noncommercial 1.0.0. Companies can evaluate Turf free for 60 days and license it for production at turf.build/licensing.

What Turf is

Governance and autonomy, the same feature.

Turf’s operating model is the Terraform® plan: every change is planned, reviewed and approved before it runs, the engine carries it out and the agent oversees it, so governance and autonomy are the same feature. Turf takes an AI factory from pools of inventory to a GPU cluster that has passed NVCRE certification, on managed Kubernetes® or bare metal, and runs day 2 on the provider’s runbook: re-certify on change, replace a failed node, upgrade, retire.

Under the plan is Turf’s new engine, built for the scale and the multi-round shape of that stack. It runs on Restate — in Restate’s words, “a lightweight runtime to turn AI agents, workflows, and backend services into durable processes”: every step of a run is journaled as it happens, so an interrupted run resumes where it stopped, a plan waiting for approval holds no process, and a rollout outlives the conversation that started it. On top, agents drive the plan over MCP, and the provider’s own control plane does the same. NVIDIA has shipped AI Cluster Runtime (AICR) 1.0 and says itself that it is “not a cluster provisioner or lifecycle management system”: you bring the cluster and the deployment tooling. Turf is what takes the cluster the rest of the way.

The plan tells the agent what is possible; the agent decides what happens.

What 1.0 stands on.

Four things, each shown rather than claimed.

01 · The engine

Plans per resource, converges in rounds

Durable per-resource state instead of one file and one lock. Layered stacks, a cluster and then the runtime on it, converge in one run and tear down in the right order. An interrupted run picks up where it stopped; a plan waiting for approval holds no process.

watch a stack converge →
02 · The operating model

Every change is a Terraform plan

Reviewed and approved before it runs. Policy with the tools you already have: the plan exports for OPA, and a hard-mandatory rule stops it before the apply. Agents drive the plan over MCP, and so does the platform you already run.

a policy the agent can’t approve its way past →
03 · The range

Three ecosystems, run as they are

Terraform providers and modules, Ansible® playbooks, Helm® charts, each as it is. Day 2 as planned Terraform actions, with lifecycle triggers and failure handling.

what each one carries →
04 · The reference architecture

Metal to Tokens

A golden recipe, one convergence loop, day 2 on the runbook: a neocloud from pools of inventory to NVCRE certification and a served model, on managed Kubernetes or bare metal. Seventeen pages, free.

read the paper →

One engine, from a laptop to the data center.

The same engine runs every stack, on a workstation or in the cluster.

Built on three ecosystems.

Nothing in this stack is written from nothing. Terraform providers and modules, Ansible playbooks and Helm charts already cover it, and Turf runs each as it is.

Terraform · providers and modules

As resources and modules

7,388 providers and 24,728 modules on the public registry reach clouds, bare-metal services, switch fabrics, storage arrays and Kubernetes. Modules package them into golden recipes.

Ansible · playbooks

As a Terraform action

4,568 collections on Ansible Galaxy cover what no API reaches: drivers, firmware, kernel settings, switch operating systems, Kubernetes on bare metal. A playbook runs as a Terraform action.

Helm · charts

As a Terraform resource

18,096 charts on Artifact Hub are how the runtime ships: the GPU Operator, NVSentinel, KAI Scheduler, Kubeflow Trainer. A chart installs as a Terraform resource.

Counts as of Metal to Tokens, October 2026. Day 2 is Terraform actions too: a lifecycle trigger runs the host playbook after a node is created, waits on a Kubernetes condition, or re-certifies after a change, and a failure halts the run or taints the resource, as the trigger says.

What’s in 1.0.

What comes next

turf setup and a Claude Code plugin, so the connect step is one command inside the agent; driving Turf from other agents over the network; provider skills that encode a provider’s own rules; and the saved-planfile round-trip, which shipped in 1.3.

Get Turf 1.0.

A download, no form, no login. Free for noncommercial use. Commercial use, including a company’s internal use, requires a license, and companies evaluate Turf under a free 60-day pilot on a standard contract.

Get started or read the terms at turf.build/licensing